Datasets:

happy8825
/

anet_sampled_sft_valid

Modalities:

Text

Formats:

Size:

Libraries:

Dataset card Data Studio Files Files and versions

xet

Community

Dataset Viewer

Auto-converted to Parquet Duplicate

Split (1)

validation · 50 rows

messages listlengths 2 2	videos listlengths 5 5
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A man receives a close and careful beard shave before the video fades to black.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] This video shows a person holding and grooming a cat. The focus is on trimming the cat's claws, with repeated sequences of holding paws and using nail clippers. There is no shaving activity and no human face shown.\n\n[2] This video features a detailed sequence of a man receiving a beard shave. It starts with applying shaving cream, followed by a close shave with a straight razor, and ends with the man cleaned up, fitting well with the query's description.\n\n[3] The video covers a man getting a haircut and styling with scissors and combs. It involves hair trimming and blow-drying but focuses on hair instead of a beard, which does not fit the query.\n\n[4] This video shows a sequence focused solely on a man getting a beard shave. The progression is clear, with the application of shaving cream and razor work leading to a clean shave, directly matching the query.\n\n[5] The video appears to be a humorous skit or challenge with various individuals performing parodic actions. It lacks any shaving or grooming activity, unrelated to the query.\n</content>\n\n<contrast>\n- [4] is better than [2] because it is entirely focused on a beard shave without distraction.\n- [2] is better than [3] because it involves facial grooming, matching the query closer than hair cutting.\n- [3] is better than [5] because it involves some grooming activity, albeit not the right kind.\n- [5] is better than [1] because it shows humans, although unrelated activities, whereas [1] is about pet grooming.\n- [1] is the least relevant as it's about grooming a cat.\n</contrast>\n\n<summary>\n1st=[4] (Best match: thorough beard shaving sequence)\n2nd=[2] (Good match: includes beard shaving but with distractions)\n3rd=[3] (Marginal match: involves grooming but of hair, not beard)\n4th=[5] (Poor match: unrelated skits involving people)\n5th=[1] (Worst match: pet grooming, no shaving or humans)\n</summary>\n</think>\n<answer> [4] > [2] > [3] > [5] > [1] </answer>", "role": "assistant" } ]	[ "activitynet/videos/v_Ac_8KQGAe0c.mp4", "activitynet/videos/v_uvnrLngXHh8.mp4", "activitynet/videos/v_hPQEkPZmNsU.mp4", "activitynet/videos/v_kq7KVmvkYPc.mp4", "activitynet/videos/v_ldjIg23Rofw.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A woman demonstrates how to prepare and enjoy rosemary vodka lemonade, showing each step from mixing the ingredients to serving and tasting the finished drink.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The video shows a woman cooking, likely preparing pasta. The sequence includes cooking onions and using a skillet, with actions centered around cooking rather than drink preparation.\n\n[2] This video displays a process involving handling various fruit slices, pouring liquids, and garnishing glasses. However, it does not focus specifically on rosemary vodka lemonade.\n\n[3] Here, a man is shown boiling pasta in a pot, with actions focused on cooking pasta rather than mixing drinks.\n\n[4] The video prominently features a woman making rosemary vodka lemonade. It shows the entire process from presenting ingredients, mixing in a bowl, pouring into glasses, and adding a rosemary garnish, followed by tasting the drink.\n\n[5] The video illustrates someone baking a cake, mixing ingredients, and putting them in the oven, unrelated to drink preparation.\n</content>\n\n<contrast>\n- Video [4] is perfect because it explicitly features making rosemary vodka lemonade with all steps clearly shown, whereas [2] shows drink preparation without specific connection to the query.\n- Video [2] shows some elements of drink preparation where [1], [3], and [5] focus on unrelated cooking activities.\n- Video [1] involves a cooking process on a stove, which doesn't match the query's focus on beverages, whereas [3] similarly involves cooking pasta.\n- Video [5] shows baking, which is entirely unrelated to the drink-making theme of the query.\n</contrast>\n\n<summary>\n1st=[4] (matches query exactly with all steps of rosemary vodka lemonade preparation)\n2nd=[2] (shows drink preparation elements but not specific to rosemary vodka lemonade)\n3rd=[1] (unrelated cooking demonstration, generally cooking pasta)\n4th=[3] (involves cooking but not relevant to the drink query)\n5th=[5] (depicts baking, entirely unrelated to the query)\n</summary>\n</think>\n<answer> [4] > [2] > [1] > [3] > [5] </answer>", "role": "assistant" } ]	[ "activitynet/videos/v_loYCYGQpGRw.mp4", "activitynet/videos/v_6lIi6-Y6q10.mp4", "activitynet/videos/v_ORL5yMppl8E.mp4", "activitynet/videos/v_qfb0Qe7s9oo.mp4", "activitynet/videos/v_KpoH4pBp9fg.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"Two men perform a synchronized and creative drumming routine on four congas, incorporating their bodies to produce various sounds and finishing with a playful gesture toward each other.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] This video shows a man playing a standard drum kit, not specifically a synchronized routine on congas or making sounds with bodies. There is no playful gesture at the end.\n\n[2] Features a single person playing two congas, lacking both the involvement of two people and the interaction described in the query.\n\n[3] Displays a solo dance performance on stage. There's no drumming or playfulness between two men.\n\n[4] Shows a man playing the saxophone, which doesn't match the conga drumming routine.\n\n[5] Depicts two men playing four congas together. The sequence shows them coordinating their actions, which aligns with the idea of a synchronized and creative routine, and includes a light final gesture that resembles playfulness.\n</content>\n\n<contrast>\n- [5] beats [1] because it matches the conga drumming setup and involves two people.\n- [5] beats [2] because it involves two performers and shows interaction, unlike the solo performance.\n- [5] beats [3] because it matches the activity of drumming rather than dancing.\n- [5] beats [4] because it involves conga drums, not saxophone.\n- [2] beats [1] and [4] because it at least involves conga drums, even if a solo.\n- [1] beats [4] because it at least involves percussion instruments.\n</contrast>\n\n<summary>\n1st=[5] (matches conga drumming, two performers, playful interaction)\n2nd=[2] (conga drumming, but solo)\n3rd=[1] (percussion instruments, single performer)\n4th=[4] (different instrument, single performer)\n5th=[3] (unrelated performance type)\n</summary>\n</think>\n<answer> [5] > [2] > [1] > [4] > [3] </answer>", "role": "assistant" } ]	[ "activitynet/videos/v_DYahQgHG1Gk.mp4", "activitynet/videos/v_TGVWaA_xTGE.mp4", "activitynet/videos/v_cNvTm72aGcM.mp4", "activitynet/videos/v_I5N4rwhRa0o.mp4", "activitynet/videos/v_UWusSRe8Whg.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A person completes a sit-up exercise routine in a gym, takes a break, and the video concludes with a feather graphic.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] Shows a group of children learning basketball drills in a gym, focused on a one-two step lay-up, unrelated to sit-ups or a feather graphic. \n\n[2] Displays a personal training session with two people kneeling and possibly exercising, but no sit-ups or feather graphic are present.\n\n[3] A person is shown performing sit-ups on a mat, and the sequence ends with a feather graphic, matching the query entirely.\n\n[4] Features an advertisement for exercise equipment, showing usage of a step machine without any sit-up exercises or feather graphics.\n\n[5] Depicts a weightlifting competition, with athletes performing lifts, which do not relate to sit-ups or a feather graphic.\n</content>\n\n<contrast>\n- [3] fully matches the query with visible sit-ups and a feather graphic, whereas [1], [2], [4], and [5] do not have these elements.\n- [2] shows a gym setting with exercise, but lacks sit-ups and a feather graphic, whereas [3] includes both.\n- [1], [4], and [5] have no elements of sit-ups or feather graphics, reducing their relevance significantly.\n</contrast>\n\n<summary>\n1st=[3] (Complete match with sit-ups and a feather graphic)\n2nd=[2] (Gym setting relevant for exercise, but lacks sit-ups and feather)\n3rd=[1] (Sports activity but unrelated to query requirements)\n4th=[5] (Sports event but unrelated to sit-ups)\n5th=[4] (Focus on exercise equipment unrelated to sit-up routine or feather graphic)\n</summary>\n</think>\n<answer> [3] > [2] > [1] > [5] > [4] </answer>", "role": "assistant" } ]	[ "activitynet/videos/v_y3Wfx-RvTpk.mp4", "activitynet/videos/v_mZYqH82delk.mp4", "activitynet/videos/v_zNpILhjhFbQ.mp4", "activitynet/videos/v_pA8QJ2ZoeBM.mp4", "activitynet/videos/v_YK5n3YTf8pI.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"Two men perform stunts involving a trampoline and a truck, jumping, lying down, and interacting with each other as part of a dynamic routine.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] Shows a sumo wrestling match with two men bowing, positioning, and engaging in the sport. No trampoline or truck is visible.\n\n[2] Features two men on a trampoline next to a red truck. They perform stunts, including jumping and interacting with each other. The setting directly involves both a trampoline and a truck.\n\n[3] Depicts weightlifting indoors. One man is lifting weights while another watches. No trampoline or truck visible.\n\n[4] Shows an athlete participating in a hammer throw event in a park setting. No trampoline or truck involved.\n\n[5] Illustrates gymnastics on bars with acrobatics. No trampoline or truck involved.\n</content>\n\n<contrast>\n- [2] is ranked higher than others because it features a trampoline, a truck, and two men performing stunts, aligning perfectly with the query.\n- [1] is in a sumo wrestling match, whereas [2] includes elements of the query.\n- [3] involves weightlifting, lacking any elements of a trampoline or truck.\n- [4] depicts a hammer throw, unrelated to trampolines or trucks.\n- [5] shows gymnastics without any involvement of a trampoline and truck.\n</contrast>\n\n<summary>\n1st=[2] (It features two men, a trampoline, and a truck, matching the query perfectly.)\n2nd=[1] (Sumo wrestling is dynamic but unrelated to trampolines or trucks.)\n3rd=[3] (Weightlifting, no trampoline or truck involved.)\n4th=[4] (Hammer throw event, unrelated to the query elements.)\n5th=[5] (Gymnastics without trampoline or truck.)\n</summary>\n</think>\n<answer> [2] > [1] > [3] > [4] > [5] </answer>", "role": "assistant" } ]	[ "activitynet/videos/v_slUvIngXl4g.mp4", "activitynet/videos/v_InavXU3iWBE.mp4", "activitynet/videos/v_m5YvKrjGtPM.mp4", "activitynet/videos/v__yk7gjnHkjs.mp4", "activitynet/videos/v_mc5XVTo5PVk.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A man splits wood on a tree stump, attempting several times to cut it before the video ends with the credits.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The video shows a person working indoors, likely plastering walls near a window. There is no wood splitting, and no tree stump is visible. This is unrelated to the query.\n\n[2] This video features a person clearing leaves in a yard using a machine and hose, and does not involve any wood splitting or tree stump. This content does not match the query.\n\n[3] The video captures a man splitting wood on a tree stump. The sequence includes lifting logs, attempting to split them, and concludes with credits. This closely matches the query.\n\n[4] A person is sawing branches on a tree stump, but there is no wood splitting or multiple attempts. It ends with credits, but the action does not match the splitting described in the query.\n\n[5] The video depicts roof repair work, including shingling and nailing. No wood splitting or tree stump is involved. It is unrelated to the query.\n</content>\n\n<contrast>\n- [3] is relevant because it directly features wood splitting on a stump, concluding with credits, whereas [4] involves sawing, not splitting.\n- [4] includes a tree stump but focuses on sawing branches, whereas [3] focuses on splitting wood, which fits the query better.\n- [1], [2], and [5] are unrelated activities, with [1] involving wall work, [2] leaf blowing, and [5] roof repair.\n</contrast>\n\n<summary>\n1st=[3] (features man splitting wood, matches query)\n2nd=[4] (includes tree stump but lacks wood splitting)\n3rd=[1] (work involves plastering, unrelated)\n4th=[2] (yard work with leaves, unrelated)\n5th=[5] (roof repairs, unrelated)\n</summary>\n</think>\n<answer> [3] > [4] > [1] > [2] > [5] </answer>", "role": "assistant" } ]	[ "activitynet/videos/v_2KYGAq5qQCI.mp4", "activitynet/videos/v_hQl2eeYX3IY.mp4", "activitynet/videos/v_yiFKPKXevOU.mp4", "activitynet/videos/v_CX5QUmM97DI.mp4", "activitynet/videos/v_9mpPAwSmUmk.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A woman explains her actions to the camera while demonstrating how to spread dirt around a tree using a bucket and a rake.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The video shows a person in a park using some kind of stilts, moving around a grassy area and sometimes on a path. There's no interaction with any tools related to gardening.\n\n[2] A person is seen using a large tool similar to a rake or broom, clearing leaves in a yard. There's no interaction with dirt or planting, nor any explicit explanation of actions to a camera.\n\n[3] A woman is demonstrating composting in a garden. She uses tools, including a small rake, and seems to be explaining the process while working with soil. This aligns with demonstrating actions and explaining to the camera in a gardening context.\n\n[4] The video shows several people on motorbikes, potentially on a motocross track. There is no demonstration involving gardening or using a bucket and rake.\n\n[5] A woman appears to be explaining a gardening task while spreading mulch or dirt around a tree using a bucket and a rake. She is engaged directly with the camera, reflecting the query requirements.\n</content>\n\n<contrast>\n- [5] is most relevant because it directly shows a woman using a bucket and rake, explaining her actions to the camera, exactly matching the query.\n- [3] is also relevant as it shows a woman explaining her actions with a rake and soil, but not around a tree specifically.\n- [2] involves outdoor work with a rake but lacks an explanatory aspect or a focus on dirt and planting around a tree.\n- [1] and [4] do not involve any gardening activities; [1] shows stilt walking and [4] is about motocross.\n</contrast>\n\n<summary>\n1st=[5] (Direct relevance: woman explains gardening tasks using a bucket and rake around a tree)\n2nd=[3] (Partial relevance: explanation related to soil and gardening with tools in a garden)\n3rd=[2] (Remote relevance: outdoor maintenance with a rake, missing explanatory and planting aspects)\n4th=[1] (Irrelevant to gardening: person on stilts in a park)\n5th=[4] (Irrelevant to gardening: motocross activities)\n</summary>\n</think>\n<answer> [5] > [3] > [2] > [1] > [4] </answer>", "role": "assistant" } ]	[ "activitynet/videos/v_NLTd7qCLT7o.mp4", "activitynet/videos/v_hQl2eeYX3IY.mp4", "activitynet/videos/v_al_769KF6Qc.mp4", "activitynet/videos/v_GNzaxjYL0TM.mp4", "activitynet/videos/v_B8WIh6PUjE0.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A man mows his lawn while carrying his child on his back, with the child calmly observing the activity.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] A child holding a lacrosse stick, playing on a deck. The child is alone and there is no adult carrying them.\n[2] A series of clips from various events, mostly involving piñatas at parties. No sign of lawn mowing or carrying a child.\n[3] A man mowing the lawn with a child on his back, exactly matching the query with the child calmly observing as they are being carried.\n[4] A person mowing a lawn, but they are alone and not carrying a child.\n[5] A group of people in a backyard engaging in casual conversations and activities, unrelated to lawn mowing or child carrying.\n</content>\n\n<contrast>\n- [3] is a perfect match because it shows a man mowing a lawn with a child on his back, whereas [4] depicts lawn mowing but without a child.\n- [4] involves lawn mowing, whereas [5] is unrelated to both lawn mowing and carrying a child.\n- [2] involves party activities unrelated to lawn mowing, whereas [1] is about a child playing, unrelated to mowing or carrying.\n</contrast>\n\n<summary>\n1st=[3] (perfect match: man mowing lawn with child on his back)\n2nd=[4] (involves lawn mowing but lacks a child)\n3rd=[5] (outdoor scene but doesn't involve mowing or child-carrying)\n4th=[1] (child playing, unrelated to mowing or carrying)\n5th=[2] (party activity, no mowing or child-carrying)\n</summary>\n</think>\n<answer> [3] > [4] > [5] > [1] > [2] </answer>", "role": "assistant" } ]	[ "activitynet/videos/v_CG-7jcSB5_c.mp4", "activitynet/videos/v_ZqkJAD0CGBU.mp4", "activitynet/videos/v_kyMsxlpBjlY.mp4", "activitynet/videos/v_L7Gt1NO1nx8.mp4", "activitynet/videos/v_DozJpP2cClI.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A newsman reports while older women sew blankets in the background.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The video shows a news reporter in a studio setting introducing a segment called \"The Needle Brigade,\" followed by scenes of older women sewing and crafting in a communal area. The transition from the reporter to the sewing activities aligns with the query.\n\n[2] This video involves various scenes of vacuum cleaners being displayed and explained, with a news reporter appearing at the end. It does not relate to the sewing or news reporting on sewing activities.\n\n[3] The frames depict people playing badminton in a backyard and later sitting on chairs, which does not align with the concept of sewing or news reporting.\n\n[4] The video begins with a news anchor in a studio, then focuses on an outdoor charity event. The broadcast and interview elements are present, but there is no sewing activity.\n\n[5] This video shows a person clearing snow from cars in a snowy environment. There is no connection to sewing or news reporting activity in this context.\n</content>\n\n<contrast>\n- [1] effectively combines news reporting with scenes of older women sewing, directly matching the query, whereas [2], [3], [4], and [5] do not.\n- [2] focuses on a completely different topic of vacuum cleaners, whereas [1] includes relevant sewing scenes.\n- [3] depicts outdoor recreational activities, lacking any news or sewing elements found in [1].\n- [4] has a news element but is centered around a charity event, lacking the specific sewing context of [1].\n- [5] presents snow shoveling, diverging from the sewing and news-centric theme in [1].\n</contrast>\n\n<summary>\n1st=[1] (Accurate depiction of a news report with sewing activities)\n2nd=[4] (Includes news reporting but lacks sewing activities)\n3rd=[2] (Features news reporter but unrelated content of vacuum cleaners)\n4th=[3] (Unrelated outdoor activity with no news or sewing component)\n5th=[5] (Irrelevant snow-clearing activity)\n</summary>\n</think>\n<answer> [1] > [4] > [2] > [3] > [5] </answer>", "role": "assistant" } ]	[ "activitynet/videos/v_kbRciA51ouY.mp4", "activitynet/videos/v_F-WmsfI8HG0.mp4", "activitynet/videos/v_DozJpP2cClI.mp4", "activitynet/videos/v_p800u2wCKbE.mp4", "activitynet/videos/v_3UbQ0UDmbbw.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A girl plays the accordion at a table, engaging with the camera as she performs.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The video shows a young girl performing a gymnastics routine on a balance beam in a gymnasium setting. There is no accordion or interaction with the camera as described in the query.\n[2] A ballet dancer in a theater setting is performing on stage. This does not involve an accordion or interaction with the camera.\n[3] A woman is performing a dance in a studio setting, but there is no accordion or interaction with the camera.\n[4] A girl is sitting at a table, playing an accordion and appears to be engaging with the camera throughout the video.\n[5] The video depicts a woman sitting and smoking indoors, with no accordion present.\n</content>\n\n<contrast>\n- [4] matches the query by featuring a girl playing an accordion and interacting with the camera, whereas [1], [2], [3], and [5] do not involve an accordion or camera engagement.\n- [1], [2], and [5] are all performances but lack any accordion playing or interaction similar to the query.\n- [3] has a dance performance but lacks both the accordion and the camera engagement described in the query.\n</contrast>\n\n<summary>\n1st=[4] (The girl is playing an accordion and engaging with the camera, matching the query description.)\n2nd=[1] (Gymnastics performance, but no accordion or camera interaction.)\n3rd=[2] (Ballet performance, lacking accordion or camera interaction.)\n4th=[3] (Dance performance, no accordion or camera interaction.)\n5th=[5] (Indoor setting, no accordion, no performance matching the query.)\n</summary>\n</think>\n<answer> [4] > [1] > [2] > [3] > [5] </answer>", "role": "assistant" } ]	[ "activitynet/videos/v_g1lqpONbOKY.mp4", "activitynet/videos/v_sNnNDCx6RRA.mp4", "activitynet/videos/v_J_SD_hhGET8.mp4", "activitynet/videos/v_-AjZCBMb4qU.mp4", "activitynet/videos/v_zuNQFkkyBNo.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"An elderly woman teaches a young girl how to knit, sharing a joyful and engaging moment together on the couch.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The video features individuals engaging in various outdoor activities, including skateboarding and scootering. They are seen both preparing for and executing tricks, primarily on a suburban street.\n\n[2] This sequence shows a person playing the piano, potentially in a recital setting, with an audience present. The focus is on the person’s hands and the intimate act of playing music.\n\n[3] An elderly woman is seen teaching a young girl how to knit. They are sitting together on a couch, sharing a joyful and engaging moment, matching the query perfectly.\n\n[4] This depicts a family decorating a Christmas tree. Individuals are seen setting up and embellishing the tree in a warmly lit living room, indicating a festive family activity.\n\n[5] A woman is peeling vegetables, specifically potatoes, in a rustic setting. The focus is on the solitary act of preparing food, without any apparent interaction or shared activity.\n</content>\n\n<contrast>\n- [3] shows a scene that directly matches the query with an elderly woman teaching a child to knit, whereas others do not feature knitting or a similar interaction.\n- [4] involves family interaction but lacks the specific one-on-one teaching element present in [3].\n- [2] and [5] lack both knitting and the intergenerational aspect; [2] focuses on solo piano playing and [5] on solitary cooking.\n- [1] shows outdoor activities unrelated to knitting or the described indoor scene.\n</contrast>\n\n<summary>\n1st=[3] (direct match to query, showing knitting lesson and joyful interaction)\n2nd=[4] (shared activity, but different context, no knitting or teaching)\n3rd=[5] (solitary, no knitting, but clear close-up)\n4th=[2] (piano, unrelated to knitting or shared teaching)\n5th=[1] (outdoor activities, least related to knitting or sitting scenario)\n</summary>\n</think>\n<answer> [3] > [4] > [5] > [2] > [1] </answer>", "role": "assistant" } ]	[ "activitynet/videos/v_DXIHbnyfNGA.mp4", "activitynet/videos/v_bCdIcqRYJvo.mp4", "activitynet/videos/v_xxXspvK4tEY.mp4", "activitynet/videos/v_dgbGv3G4k5M.mp4", "activitynet/videos/v_UzCeMCr0lM8.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A woman demonstrates showcasing a doll's hair after removing its curlers, highlighting its appearance from different angles.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] This is about grooming a poodle. A woman interacts with a dog who appears unruly, and there is no doll or demonstration of hair being styled.\n\n[2] A woman is seen demonstrating a doll's hair with curlers, removing them, and showcasing the hair from different angles. This closely matches the query.\n\n[3] The video shows a woman styling her own hair, creating a hairdo. It does not involve a doll or removing curlers.\n\n[4] A man styles a mannequin's hair, cutting and shaping it. It is not related to a woman demonstrating a doll's hair with curlers.\n\n[5] A woman styles her own hair with steps on how to achieve a specific look. There is no doll involved, or any curlers being removed.\n</content>\n\n<contrast>\n- [2] directly shows a woman removing curlers from a doll's hair, whereas [1], [3], [4], and [5] do not involve a doll or curlers.\n- [3] and [5] focus on a woman styling her own hair, whereas the query specifies a doll.\n- [4] involves a man styling a mannequin’s hair, which does not match the query about a woman's demonstration.\n</contrast>\n\n<summary>\n1st=[2] (Directly matches the query by featuring a doll’s hair with curlers being removed by a woman)\n2nd=[4] (Involves hair styling on a mannequin which is closer than self-styling)\n3rd=[3] (Demonstrates hair styling but on a person, not a doll)\n4th=[5] (Similar to [3] but less action and interaction visible)\n5th=[1] (Not relevant to dolls or hair demonstration)\n</summary>\n</think>\n<answer> [2] > [4] > [3] > [5] > [1] </answer>", "role": "assistant" } ]	[ "activitynet/videos/v_W04glwJ_IYI.mp4", "activitynet/videos/v_RQwgREUcreI.mp4", "activitynet/videos/v_4-_ZQGwppfI.mp4", "activitynet/videos/v_JU_o9ZtH-VM.mp4", "activitynet/videos/v_padyJHC5Y5Y.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A man demonstrates and explains proper knife sharpening techniques using various tools and a sharpening board.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] This video shows a person assembling a tire with tools, not related to knife sharpening.\n\n[2] A person is demonstrating exercises on a bench, unrelated to knife sharpening.\n\n[3] The frames depict a person handling a sharpening stone and a tool resembling a knife, directly related to knife sharpening techniques.\n\n[4] This sequence involves a person working on wood painting and finishing, unrelated to knife sharpening.\n\n[5] The video depicts a person installing flooring, not connected to knife sharpening.\n</content>\n\n<contrast>\n- [3] is directly related to knife sharpening, whereas [1], [2], [4], and [5] are not.\n- [1] and [5] involve other types of tool usage, whereas [3] specifically focuses on sharpening.\n- [2] and [4] are about entirely different activities (exercise and woodworking) compared to [3].\n</contrast>\n\n<summary>\n1st=[3] (demonstrates knife sharpening with applicable tools)\n2nd=[1] (shows tool usage, though not sharpening)\n3rd=[5] (involves tool handling, but irrelevant domain)\n4th=[2] (completely unrelated exercise demonstration)\n5th=[4] (focused on woodworking, not relevant)\n</summary>\n</think>\n<answer> [3] > [1] > [5] > [2] > [4] </answer>", "role": "assistant" } ]	[ "activitynet/videos/v_4r0qYLCucEI.mp4", "activitynet/videos/v_FujUrbL1qc0.mp4", "activitynet/videos/v_Anvhv9FFv1Q.mp4", "activitynet/videos/v__JE5T2RKZvo.mp4", "activitynet/videos/v_ZIBWRRBft8g.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"An exciting indoor soccer match unfolds as teams in yellow and red compete fiercely, exchanging goals and dramatic saves before the yellow team ultimately wins, celebrates, and is awarded the trophy amid cheers from the crowd.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The frames show an indoor soccer match between teams in yellow and red uniforms. The sequence includes goal exchanges, dramatic saves, and ends with the yellow team celebrating with a trophy presentation, fitting the query closely.\n\n[2] This sequence includes frames of an indoor soccer match with teams in red and yellow. It captures action scenes with players in motion, although there's no evidence of a trophy celebration or clear indication of a match ending.\n\n[3] This video consists of news reporters and interviews with some indoor soccer match footage, but lacks the decisive match elements, goal exchanges, or celebrations depicted in the query.\n\n[4] The frames capture a casual indoor sports setting with students, possibly gym class or informal play, not the intense match or trophy award described in the query.\n\n[5] Displays various futsal matches with teams wearing different colored uniforms. The sequence contains general match footage without specific celebratory or trophy-related scenes.\n\n</content>\n\n<contrast>\n- [1] has the complete narrative of goals, saves, celebration, and trophy, whereas [2] contains only match footage and lacks the celebratory ending.\n- [2] includes relevant soccer action more than [3], which mixes news and sparse match footage.\n- [3] involves some match elements but lacks coherence; compared to [4], it's slightly more relevant with sports content.\n- [4] lacks structured match scenes, whereas [5] shows organized matches but without the ending celebration.\n- [5] has more match-related content than [3] or [4], though it misses the celebratory scene like [2].\n\n</contrast>\n\n<summary>\n1st=[1] (complete match narrative with celebration)\n2nd=[2] (match footage but lacks celebration)\n3rd=[5] (organized matches but no trophy scene)\n4th=[3] (news and sparse match content)\n5th=[4] (casual sports setting, not an organized match)\n\n</summary>\n</think>\n<answer> [1] > [2] > [5] > [3] > [4] </answer>", "role": "assistant" } ]	[ "activitynet/videos/v_ynda_PQe_CM.mp4", "activitynet/videos/v_XxMzH75Bhr8.mp4", "activitynet/videos/v_Mo9-1cuMVtQ.mp4", "activitynet/videos/v_bWYIL4KG8kM.mp4", "activitynet/videos/v_bMDyOXygNPM.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A girl methodically rides a yellow lawn mower back and forth across a large grassy yard, carefully making turns to cover the entire area.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The sequence shows a person mowing grass with a manual mower near tents in a grassy area. There is no indication of a girl or a yellow lawn mower.\n\n[2] This video displays a group of people in a backyard engaged in leisure activities, such as sitting and playing badminton. There is no lawn mowing activity visible.\n\n[3] This sequence depicts a person using stilts in a grassy park area, featuring a fountain and pathways. It doesn't include a lawn mower or a girl mowing grass.\n\n[4] The video clearly shows a girl riding a yellow lawn mower back and forth across a large grassy yard, making turns to cover the area comprehensively.\n\n[5] This compilation includes people in costume participating in a croquet event on a college campus. There is no mowing activity or lawn mower visible.\n</content>\n\n<contrast>\n- [4] matches the query as it includes a girl on a yellow lawn mower covering a grassy area, whereas [1] involves manual mowing by a different individual.\n- [4] displays the exact activity described in the query in a close-up and coherent manner, while [2], [3], and [5] involve unrelated activities.\n- [1] does show mowing, but it doesn’t match the equipment or specific action described in the query.\n- [2], [3], and [5] lack any lawn mowing activity and are not relevant to the query.\n</contrast>\n\n<summary>\n1st=[4] (Shows a girl using a yellow lawn mower as described)\n2nd=[1] (Related but involves a manual mower and different scenario)\n3rd=[2] (Unrelated; backyard gathering)\n4th=[3] (Unrelated; person on stilts in a park)\n5th=[5] (Unrelated; croquet event on campus)\n</summary>\n</think>\n<answer> [4] > [1] > [2] > [3] > [5] </answer>", "role": "assistant" } ]	[ "activitynet/videos/v_rjgu0ucCW7U.mp4", "activitynet/videos/v_DozJpP2cClI.mp4", "activitynet/videos/v_NLTd7qCLT7o.mp4", "activitynet/videos/v_Vcxv3nyZpSs.mp4", "activitynet/videos/v_avxSMcGBdG8.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A man demonstrates how to clean a pair of dirty Nike sneakers using water, cleaning solution, and a brush.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The video demonstrates roof repair using a product called EternaBond. It involves cleaning, preparing, and applying adhesive tape, unrelated to sneaker cleaning.\n\n[2] The video shows a demonstration of cleaning Nike sneakers. It includes using a cleaning solution, brush, and water, with a focus on restoring the shoes' appearance.\n\n[3] This video is about cleaning a bathroom sink, bathtub, and toilet using spray and cloths, not related to sneaker cleaning.\n\n[4] The video shows handwashing in a sink using soap, water, and paper towels, unrelated to sneaker cleaning.\n\n[5] This video shows a person washing a small dog with soap and water, which is unrelated to sneaker cleaning.\n</content>\n\n<contrast>\n- [2] is directly related to sneaker cleaning with a clear demonstration involving Nike sneakers, whereas [1], [3], [4], and [5] focus on different cleaning tasks.\n- [1] is about roof repair, which is entirely different from the sneaker cleaning in [2].\n- [3] and [4] involve bathroom cleaning and hand washing, which do not relate to sneaker cleaning.\n- [5] focuses on washing a dog, which does not pertain to sneaker cleaning like in [2].\n</contrast>\n\n<summary>\n1st=[2] (Direct demonstration of cleaning Nike sneakers using appropriate tools)\n2nd=[1] (Involves cleaning, but unrelated to sneakers)\n3rd=[3] (Cleans objects, but unrelated to sneaker cleaning)\n4th=[4] (Involves cleaning via hand washing, unrelated to sneakers)\n5th=[5] (Involves washing a pet, unrelated to sneakers)\n</summary>\n</think>\n<answer> [2] > [1] > [3] > [4] > [5] </answer>", "role": "assistant" } ]	[ "activitynet/videos/v_VcyO9tZ3NkY.mp4", "activitynet/videos/v_WE9Md637nbc.mp4", "activitynet/videos/v_eMB9L-LDK1E.mp4", "activitynet/videos/v_f6Nhpyq1XdU.mp4", "activitynet/videos/v_y76trNI6U0Y.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A woman brushes her teeth at the bathroom sink while being filmed by someone behind her.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] This video shows two women in a bathroom, focused on using mouthwash rather than brushing teeth. It does not match the query.\n[2] A child is bathing a dog in what appears to be a shower or tub area. This is unrelated to teeth brushing.\n[3] A man is using an electric razor to shave in front of a bathroom mirror, which is unrelated to the query.\n[4] A woman and a child are washing dishes at a kitchen sink, not related to teeth brushing or a bathroom setting.\n[5] This video features a woman brushing her teeth in front of a bathroom mirror, with someone likely filming her from behind, closely matching the query.\n</content>\n\n<contrast>\n- [5] is chosen first because it directly shows a woman brushing her teeth in a bathroom while being filmed. \n- [1] is more related than others as it takes place in a bathroom but involves mouthwash use, not brushing teeth.\n- [3] involves bathroom-related activity but is focused on shaving, not brushing teeth.\n- [4] involves sink-related activity but is about washing dishes, not brushing teeth or filmed in a bathroom setting.\n- [2] involves a bathing activity but not in a bathroom sink setting and is unrelated to dental hygiene.\n</contrast>\n\n<summary>\n1st=[5] (directly matches the query with teeth brushing in a bathroom)\n2nd=[1] (loosely related due to bathroom setting and oral hygiene)\n3rd=[3] (bathroom setting but unrelated to teeth brushing)\n4th=[4] (sink activity but unrelated to bathroom or teeth brushing)\n5th=[2] (unrelated activity, not bathroom or dental hygiene)\n</summary>\n</think>\n<answer> [5] > [1] > [3] > [4] > [2] </answer>", "role": "assistant" } ]	[ "activitynet/videos/v_I-vi5EpjrFI.mp4", "activitynet/videos/v_O9qI26TB3R0.mp4", "activitynet/videos/v_ze6Bu0bcNbg.mp4", "activitynet/videos/v_XbSQ_7vh7yc.mp4", "activitynet/videos/v_Zts8FynhoJs.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A man enthusiastically demonstrates his strength through arm wrestling and weight lifting, ultimately celebrating his victory in a lively manner.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] This video showcases a group cardio workout class on stationary bikes, with multiple participants pedaling together. There is no evidence of arm wrestling or weight lifting.\n\n[2] The video depicts a wrestling match in a large arena. A muscular man is seen interacting with opponents and the audience, displaying physical strength but not involving arm wrestling or explicit weight lifting.\n\n[3] The video shows a man plastering walls inside a room, demonstrating construction skills. It has no correlation with arm wrestling, weight lifting, or celebratory moments.\n\n[4] This video captures a man engaged in arm wrestling and weight lifting exercises. There are sequences showing arm wrestling matches and the use of dumbbells, potentially culminating in a celebratory or successful moment.\n\n[5] The video features a demonstration of a fitness machine with a focus on settings and usage. It includes scenes with a person exercising but lacks arm wrestling or direct weight lifting.\n</content>\n\n<contrast>\n- [4] ranks higher than others because it directly includes both arm wrestling and weight lifting, aligning well with the query.\n- [2] provides a demonstration of strength in a wrestling context, but lacks arm wrestling and weight lifting, making it less relevant than [4].\n- [1] involves a fitness setting but focuses on group cycling without the strength demonstration or celebratory context required by the query.\n- [5] involves exercise equipment but lacks the specific actions of arm wrestling and direct weight lifting.\n- [3] is unrelated as it involves plastering and construction work, not matching the query focus on strength and celebration.\n</contrast>\n\n<summary>\n1st=[4] (Directly related: arm wrestling and weight lifting shown)\n2nd=[2] (Displays strength in wrestling but lacks specific query actions)\n3rd=[1] (Involves fitness but focuses on cycling, not strength demonstration)\n4th=[5] (Exercise equipment demo, lacks arm wrestling/weight lifting)\n5th=[3] (Unrelated construction work)\n</summary>\n</think>\n<answer> [4] > [2] > [1] > [5] > [3] </answer>", "role": "assistant" } ]	[ "activitynet/videos/v_Q5kU3DUXXZg.mp4", "activitynet/videos/v_qcaaF0SqVss.mp4", "activitynet/videos/v_hhN1647pP88.mp4", "activitynet/videos/v_ecVUPpBNpdA.mp4", "activitynet/videos/v_WElK913B33U.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A woman demonstrates how to apply aging cream to her face, provides information about its weekly benefits, and encourages viewers to visit the product's website for more details.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] A woman is in a kitchen setting demonstrating knife sharpening techniques. No mention of face cream or skincare benefits.\n\n[2] A close-up tutorial on shaving a man's face with a straight razor. No indication of skincare application or discussion of benefits.\n\n[3] This is a tutorial for inserting contact lenses, focusing on eye appearance rather than skincare.\n\n[4] Similar to [2], another tutorial on shaving a man's face with a straight razor. Still not relevant to skincare or face cream benefits.\n\n[5] This video shows a woman applying a cream to her face. It includes text detailing its usage, the benefits of a retinol product, and advice to visit a webpage for more details, directly matching the query.\n</content>\n\n<contrast>\n- [5] is perfect because it directly illustrates a woman applying face cream with detailed skincare benefits and website information, whereas others do not cover face cream.\n- [1] involves knife sharpening, irrelevant to skincare, whereas [3] and [4] focus on shaving or eye makeup.\n- [2] and [4] are only about shaving, having no alignment with applying or detailing skincare creams.\n- [3] deals with contact lenses, lacking any connection to skincare or cream benefits.\n</contrast>\n\n<summary>\n1st=[5] (Directly illustrates skincare application and benefits).\n2nd=[3] (Shows a personal care product, though unrelated to face cream).\n3rd=[2] (Covers personal care, but focuses on shaving, not skincare).\n4th=[4] (Same reason as [2], but less clear on skincare elements).\n5th=[1] (Completely irrelevant to skincare).\n</summary>\n</think>\n<answer> [5] > [3] > [2] > [4] > [1] </answer>", "role": "assistant" } ]	[ "activitynet/videos/v_1SQAnbh_lcM.mp4", "activitynet/videos/v_kq7KVmvkYPc.mp4", "activitynet/videos/v_IJ2m8Bi3LVE.mp4", "activitynet/videos/v_uvnrLngXHh8.mp4", "activitynet/videos/v_bQVMoeBK7XU.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A woman is seen swimming, floating, and performing various moves in the water, including using a boogie board, as several clips showcase her aquatic activities.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The images display a woman engaging in a track and field long jump event. She is seen preparing, jumping, and landing, which is not related to swimming or aquatic activities.\n\n[2] This sequence shows various swimming and floating activities in a pool, with a focus on underwater swimming techniques. The woman is visible swimming and performing movements in the water, with instructional elements present.\n\n[3] The images depict scuba diving activities in a pool. Individuals are using scuba gear and interacting underwater, which involves water activities but not specifically swimming or floating as described in the query.\n\n[4] These frames feature a news report about a wakeboarding invention. The scenes involve a person wakeboarding, being pulled by a device on water, which is not aligned with the query of a woman swimming or using a boogie board.\n\n[5] The video sequence captures a woman swimming in a pool, performing moves, and using a swimming aid resembling a mermaid tail. This aligns with the water activities detailed in the query, focusing on diverse swimming movements.\n</content>\n\n<contrast>\n- [2] ranks higher than others because it shows the woman swimming and performing various water movements as specified in the query, with clips showcasing pool activities clearly.\n- [5] is placed next as it also displays a woman performing swimming and aquatic movements, enhancing relevance with pool scenes and swimming aids.\n- [3] follows since it involves underwater activities, but is focused on scuba diving, which is less directly relevant than swimming.\n- [4] is below [3] because it involves wakeboarding, which differs significantly from the swimming and floating activities described in the query.\n- [1] is last as it depicts track and field activities, which are unrelated to any aquatic setting or swimming.\n</contrast>\n\n<summary>\n1st=[2] (Clear depiction of swimming and aquatic movements relevant to the query)\n2nd=[5] (Visible swimming activities and use of pool, closely related to query)\n3rd=[3] (Underwater activities but focused on scuba diving rather than swimming)\n4th=[4] (Wakeboarding activities, different from swimming/floating focus)\n5th=[1] (Track and field activities, unrelated to query)\n</summary>\n</think>\n<answer> [2] > [5] > [3] > [4] > [1] </answer>", "role": "assistant" } ]	[ "activitynet/videos/v_fruimvo90vA.mp4", "activitynet/videos/v_kFmOtf_Ew1w.mp4", "activitynet/videos/v_BahcBS2r0es.mp4", "activitynet/videos/v_wdsZn1Yu_5M.mp4", "activitynet/videos/v_XC6tvSBS0PA.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A group of girls and boys gather in a room, pose for photos together, and play a game of foosball.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] This sequence takes place in a gymnasium with groups of people, mostly young, interacting and playing sports. It lacks any clear foosball action or photo posing.\n[2] Office surveillance showing people working and meeting, not consistent with the query involving foosball or casual gathering.\n[3] Shows a group of people around a foosball table, posing with medals, and playing the game. This aligns well with a casual group gathering and foosball activity.\n[4] Depicts a one-on-one fitness instruction session with no connection to group activities or foosball.\n[5] Focuses on close-ups of a foosball table in action, showing gameplay but without context of a group or posing.\n</content>\n\n<contrast>\n- [3] covers both group gathering and foosball gameplay well, whereas [1] only partially aligns with the group gathering aspect.\n- [5] shows foosball clearly, but lacks the group and posing elements present in [3].\n- [2] has no relevant content to group gatherings or foosball.\n- [4] is focused on fitness instruction, unrelated to the query elements.\n</contrast>\n\n<summary>\n1st=[3] (covers group gathering, posing, and foosball)\n2nd=[5] (shows foosball gameplay well)\n3rd=[1] (some group interaction, no foosball or posing)\n4th=[2] (unrelated office environment)\n5th=[4] (unrelated fitness session)\n</summary>\n</think>\n<answer> [3] > [5] > [1] > [2] > [4] </answer>", "role": "assistant" } ]	[ "activitynet/videos/v_bWYIL4KG8kM.mp4", "activitynet/videos/v_Po6rrhjsS8g.mp4", "activitynet/videos/v_mnv-Qz4QDfU.mp4", "activitynet/videos/v_F_DHgp5hERs.mp4", "activitynet/videos/v_HJZj6z1dyT4.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A man demonstrates and explains the process of removing and taking apart a roof using specific tools.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The video shows various stages of a concrete roof repair. It includes footage of a person applying a coating onto a roof surface, using long-handled tools. However, it lacks a clear narrative or demonstration focused on removing or taking apart the roof.\n\n[2] This video features a man explaining and demonstrating roof shingle removal. It shows him using specific tools to take apart roof shingles, fitting the query of demonstrating and explaining the roof removal process effectively.\n\n[3] The video focuses on a person doing welding work inside a garage. It does not relate to roof removal or explanation, as it purely involves metalwork activities.\n\n[4] The content involves assembling a bicycle, detailing steps to install parts and accessories. This is unrelated to roof removal or demonstration of related processes.\n\n[5] The video shows a person preparing and cooking in an outdoor setting. The actions revolve around food preparation, which is not relevant to the query's focus on roof removal.\n</content>\n\n<contrast>\n- [2] is prioritized over [1] because it shows explicit step-by-step roof shingle removal with tool demonstration, whereas [1] focuses on applying roof coatings with less emphasis on removal.\n- [1] is ranked higher than [3], [4], and [5] as it involves roof work, though not focused on removal. [3] involves welding, while [4] and [5] deal with unrelated activities (bicycle assembly and cooking).\n- [3] is above [4] because welding is closer to construction than bicycle assembly. [4] and [5] are similar in irrelevance but [4] is more process-oriented related to assembly.\n</contrast>\n\n<summary>\n1st=[2] (demonstrates and explains roof shingle removal clearly)\n2nd=[1] (focus on roof work but not removal; applying coatings)\n3rd=[3] (welding activity, not related but closer to construction)\n4th=[4] (bicycle assembly, unrelated to roof work)\n5th=[5] (cooking, entirely unrelated to roof and tool use)\n</summary>\n</think>\n<answer> [2] > [1] > [3] > [4] > [5] </answer>", "role": "assistant" } ]	[ "activitynet/videos/v_ir1jpnsK5p8.mp4", "activitynet/videos/v_bX-PSFlwV5s.mp4", "activitynet/videos/v_5Bo0gFXxDQk.mp4", "activitynet/videos/v_IbYlWVA8Dh8.mp4", "activitynet/videos/v_28P5f3p_32g.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A woman sharpens a knife behind a counter, pouring water on the sharpener and methodically preparing the blade.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] A woman is peeling a fruit, possibly a potato, but there's no evidence of sharpening a knife or using a sharpener.\n[2] A person is washing a dog, no knife sharpening or related actions depicted.\n[3] A chef is preparing food, possibly using a knife, but there's no specific focus on sharpening knives or pouring water.\n[4] A person is peeling a potato using a peeler, no knife sharpening or related activities are visible.\n[5] A woman demonstrates sharpening a knife using a stone and methodically prepares the blade, consistent with pouring water on the sharpener.\n</content>\n\n<contrast>\n- [5] is directly relevant as it shows a woman sharpening a knife with a stone whereas [1], [2], [3], and [4] do not show knife sharpening.\n- [1] shows a woman involved in a kitchen activity but lacks knife sharpening, whereas [5] includes the complete process.\n- [3] involves food preparation which may involve knives, but it doesn’t focus on sharpening like [5].\n- [4] focuses on peeling rather than sharpening as shown in [5].\n</contrast>\n\n<summary>\n1st=[5] (relevant activity of sharpening knife with water)\n2nd=[3] (involves kitchen knife but not sharpening focus)\n3rd=[1] (kitchen setting but no sharpening)\n4th=[4] (peeling activity, irrelevant to sharpening)\n5th=[2] (unrelated pet grooming activity)\n</summary>\n</think>\n<answer> [5] > [3] > [1] > [4] > [2] </answer>", "role": "assistant" } ]	[ "activitynet/videos/v_UzCeMCr0lM8.mp4", "activitynet/videos/v_y76trNI6U0Y.mp4", "activitynet/videos/v_VpfLfFtu0d8.mp4", "activitynet/videos/v_IvkpJMxlRKc.mp4", "activitynet/videos/v_1SQAnbh_lcM.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"The video demonstrates unboxing, assembling, charging, and using a Phillips three-headed razor, followed by a shaving demonstration and a return of the razor to its charger.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] This video is about a knitting and crafting session at a library. The clips show people knitting, reading knitting pattern books, and engaging in a group activity. There is no content related to razors or shaving.\n\n[2] This video involves a maintenance task, likely related to sealing and repairing a roof or an appliance part. It shows tools, sealing compound, and surfaces but has no relation to razors or shaving.\n\n[3] This video clearly shows the unboxing of a Phillips three-headed razor, followed by its use for shaving and returning to the charger. It covers all the elements mentioned in the query, including close-ups and sequential action.\n\n[4] This video demonstrates the use of a fitness machine. It shows a person using an elliptical trainer and related digital metrics. There is no content related to razors or shaving.\n\n[5] This video depicts a professional shaving service using straight razors and brushes. There are no Phillips razors or context of unboxing and charging shown.\n</content>\n\n<contrast>\n- [3] is relevant because it covers unboxing, charging, using, and returning the Phillips razor, whereas [5] focuses on professional shaving techniques without demonstrating the specific razor or actions of unboxing and charging.\n- [5] shows shaving but lacks the specific Phillips razor context of unboxing and charging, whereas [1], [2], and [4] are not related to razors or shaving at all.\n- [4] involves fitness equipment not related to the query, whereas [5] contains shaving-related content.\n</contrast>\n\n<summary>\n1st=[3] (Directly matches the query with relevance in razor unboxing, assembling, charging, and shaving demonstration).\n2nd=[5] (Related to shaving but lacks coverage of the specific razor and actions like unboxing and charging).\n3rd=[1] (No relevance to razors or shaving; focuses on knitting).\n4th=[2] (No relevance; focused on repair work).\n5th=[4] (No relevance; focused on exercise equipment).\n</summary>\n</think>\n<answer> [3] > [5] > [1] > [2] > [4] </answer>", "role": "assistant" } ]	[ "activitynet/videos/v_sRMrtzg0nwU.mp4", "activitynet/videos/v_PzAjZbCEbLU.mp4", "activitynet/videos/v_ze6Bu0bcNbg.mp4", "activitynet/videos/v_8874vqLm958.mp4", "activitynet/videos/v_uvnrLngXHh8.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A man irons a shirt, but afterwards appears to be in pain as he covers his ears and looks away.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] A man is seen ironing a shirt using an iron. Later, he appears to be in discomfort, covering his ears and turning his head away.\n[2] A man uses an electric shaver. Sequence includes showing the shaver, shaving, and examining the result in the mirror.\n[3] This features an Olympic gymnastics competition, showing athletes performing on pommel horse, crowds cheering, and medal presentations.\n[4] A man is plastering a wall, spreading plaster in an even manner across the surface.\n[5] Two people are baking cookies, shown preparing dough, placing it in the oven, and engaging in a conversation.\n\n</content>\n\n<contrast>\n- [1] is directly relevant to the query with a man ironing a shirt and then appearing to be in pain, whereas [2] involves shaving, which is off-topic.\n- [1] features the act of ironing and the subsequent pain, fitting the query perfectly, whereas [3] focuses on gymnastics and medals.\n- [1] shows continuity from ironing to discomfort, matching the query, whereas [4] involves plastering a wall, unrelated to the ironing and pain theme.\n- [1] aligns with the ironing and discomfort theme, whereas [5] involves a baking process, mismatching the query.\n- [2] is primarily about shaving, [3] about gymnastics, [4] about wall plastering, and [5] about baking, all less relevant than [1].\n\n</contrast>\n\n<summary>\n1st=[1] (Shows a man ironing a shirt then in pain, perfect match for the query)\n2nd=[2] (Although about shaving, depicts a specific personal grooming activity)\n3rd=[5] (Involves an indoor activity, albeit unrelated to ironing)\n4th=[4] (Shows individual work, but unrelated to ironing or discomfort)\n5th=[3] (Sports event, no relation to the query)\n\n</summary>\n</think>\n<answer> [1] > [2] > [5] > [4] > [3] </answer>", "role": "assistant" } ]	[ "activitynet/videos/v_RNrxxPOyHo4.mp4", "activitynet/videos/v_ze6Bu0bcNbg.mp4", "activitynet/videos/v_Le6fNx5IWRI.mp4", "activitynet/videos/v_hhN1647pP88.mp4", "activitynet/videos/v_3OGD7if0zs8.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A young boy enthusiastically throws darts at a dartboard and the camera shows the results of his throws.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] A young boy is shown throwing darts at a dartboard. The sequence includes close-up shots of the dartboard with darts stuck in it, and features the boy actively moving towards or away from the setup, effectively highlighting both the action and the results.\n\n[2] A man in a snowy outdoor setting prepares to shoot arrows at an archery target. He is holding a bow, and there are several frames focused on the archery equipment and the target.\n\n[3] Several people are seen in a bar setting with dartboards visible on the walls. However, there is no clear focus on a single player or results shown on the dartboards.\n\n[4] Two people are playing squash in an indoor court. The actions involve swinging rackets and moving across the court, unrelated to dart throwing.\n\n[5] A young boy enthusiastically throws darts at a dartboard, with several frames dedicated to close-ups of the dartboard showing darts successfully hitting the target. The sequence is centered on the boy's actions and the results of the throws.\n</content>\n\n<contrast>\n- [5] is ranked higher than [1] because it has more emphasis on both the boy's action of throwing darts and the results shown on the dartboard.\n- [1] ranks above [3] as it focuses on a young boy throwing and hitting darts, whereas [3] lacks a clear demonstration of the action or results.\n- [3] is rated higher than [2] because it at least includes dartboards, while [2] features an entirely different activity—archery.\n- [2] beats [4] since it involves target shooting, which is somewhat related to darts, unlike the squash game depicted in [4].\n</contrast>\n\n<summary>\n1st=[5] (Fully shows the boy throwing darts and the results)\n2nd=[1] (Shows both the boy throwing and dartboard results, but less complete than [5])\n3rd=[3] (Dartboards are present, but less clear action and results)\n4th=[2] (Unrelated archery action, but involves target shooting)\n5th=[4] (Completely unrelated squash game)\n</summary>\n</think>\n<answer> [5] > [1] > [3] > [2] > [4] </answer>", "role": "assistant" } ]	[ "activitynet/videos/v_JKVOi9DE8Hs.mp4", "activitynet/videos/v_4Xvn1xXvYdU.mp4", "activitynet/videos/v_smGijLg8Cho.mp4", "activitynet/videos/v_FayB7ZswnAw.mp4", "activitynet/videos/v_UJiQEkByWcM.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A man flies a wind kite on a grassy field while dogs run around nearby.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The scene involves a group of people playing a sport on a grassy field. There is no evidence of kite flying or the presence of dogs.\n\n[2] A man is seen flying a wind kite on a grassy field with a dog running around nearby. This matches the query closely, including the presence of the kite and the dog.\n\n[3] The video shows people in orange shirts standing on a sandy field, likely a baseball or softball game. There is no kite flying or dogs present.\n\n[4] The video shows people playing in a backyard and sitting around, but there is no evidence of kite flying or dogs.\n\n[5] Two people engage with what appears to be a suspended soccer ball on a field, involving ropes and poles, but no kites or dogs are seen.\n</content>\n\n<contrast>\n- [2] matches the query with both kite flying and a dog running nearby, whereas [1] has no kite flying or dogs.\n- [3] features a field-related activity, but there's no kite flying or dogs, unlike [2].\n- [4] involves casual activity without kite flying or dogs, setting it apart from the close match in [2].\n- [5] involves activity with ropes on a field but no kite or dogs, unlike [2] that shows both.\n</contrast>\n\n<summary>\n1st=[2] (Contains kite flying and a dog on a grassy field, matching all aspects of the query.)\n2nd=[1] (Involves field activity, but no kite or dogs are present.)\n3rd=[5] (Shows field activity with ropes and poles, but no kite or dogs.)\n4th=[3] (Features a different field activity, not involving a kite or dogs.)\n5th=[4] (Shows casual backyard activity, unrelated to query elements.)\n</summary>\n</think>\n<answer> [2] > [1] > [5] > [3] > [4] </answer>", "role": "assistant" } ]	[ "activitynet/videos/v_wUg9S2xx850.mp4", "activitynet/videos/v__O6ONcIPR_s.mp4", "activitynet/videos/v_0jBwj0bfZ3Y.mp4", "activitynet/videos/v_DozJpP2cClI.mp4", "activitynet/videos/v_OtcRXishIxw.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A hockey game takes a dramatic turn as players collide and one suffers a serious head injury on the ice.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] This video shows various hockey players on the ice, with no apparent collision or injury being central to the action. There's focus on a goalie and other non-injury related activities.\n[2] This video depicts a person assembling and flying a kite in an outdoor setting. It has no relation to hockey or any injuries.\n[3] This video shows a field sport with players, coaches, and spectators, likely a hurling match. There is no evidence of a hockey game or injuries.\n[4] This video involves people playing mini pool, with no relation to a hockey game or any dramatic turn related to injuries.\n[5] This video sequence includes multiple instances of hockey players colliding, falling, and lying injured on the ice, closely aligning with the query about a dramatic turn and a serious head injury.\n\n</content>\n\n<contrast>\n- [5] shows clear collisions and injuries during a hockey game, whereas [1] shows general hockey gameplay without such incidents.\n- [2] is unrelated to the query with its focus on kite flying.\n- [3] involves a different sport with no injuries resembling the described scenario.\n- [4] focuses on a pool game and lacks any connection to hockey or injuries.\n\n</contrast>\n\n<summary>\n1st=[5] (Clearly depicts hockey players colliding and suffering injuries, matching the query.)\n2nd=[1] (Involves hockey, but lacks the dramatic injury aspect.)\n3rd=[3] (Another sport with players, but no hockey or injury relevance.)\n4th=[4] (No relation to hockey or injury; involves playing pool.)\n5th=[2] (Completely unrelated, focuses on a kite activity.) \n\n</summary>\n</think>\n<answer> [5] > [1] > [3] > [4] > [2] </answer>", "role": "assistant" } ]	[ "activitynet/videos/v_wZ-teiWX4mg.mp4", "activitynet/videos/v_AH-3mF6wWhU.mp4", "activitynet/videos/v_y4PXBhxpZHk.mp4", "activitynet/videos/v_TN8tBV7CY6U.mp4", "activitynet/videos/v_CEEJ1gzdmAk.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A group of girls play dodge ball outside, with one girl getting hit and leaving the game to stand by the clothesline.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The images display a cheer routine taking place indoors on a basketball court. Two girls are performing, and there is no evidence of a dodgeball game or a clothesline.\n[2] The sequence shows a group of girls playing a ball game outdoors in a yard with a brick wall and visible clothesline. One girl appears to get hit and stands by the clothesline.\n[3] This series shows women exercising on a lawn using ropes. There is no indication of a dodgeball game or clothesline.\n[4] The frames depict a gymnastics routine on a balance beam inside a gymnasium. There is no outdoor activity, dodgeball, or clothesline visible.\n[5] The scene shows children and adults outdoors with toy bats and piñatas. The setting is a paved area, without a dodgeball game or clothesline.\n</content>\n\n<contrast>\n- [2] is ranked higher because it directly shows a group of girls playing a ball game outdoors, matching the query details closely, including the presence of a clothesline.\n- [1] is not relevant as it features a cheer routine indoors, whereas the query specifies a ball game outside.\n- [3] involves an exercise routine, not a dodgeball game or outdoor setting with a clothesline.\n- [4] shows an indoor gymnastics routine, unrelated to the query's outdoor dodgeball game context.\n- [5] includes an outdoor activity but involves piñatas, not dodgeball, and lacks a clothesline context.\n</contrast>\n\n<summary>\n1st=[2] (Matches closely with girls playing a ball game outside by a clothesline.)\n2nd=[5] (Outdoor activity but involves piñatas, not dodgeball and lacks clothesline context.)\n3rd=[1] (Indoor cheer routine, not a dodgeball game outside.)\n4th=[3] (Exercise routine outdoors, lacks dodgeball or clothesline elements.)\n5th=[4] (Indoor gymnastics routine, unrelated to outdoor dodgeball game query.)\n</summary>\n</think>\n<answer> [2] > [5] > [1] > [3] > [4] </answer>", "role": "assistant" } ]	[ "activitynet/videos/v_8w9pf_Ecm_U.mp4", "activitynet/videos/v_PjSOjB7qhcA.mp4", "activitynet/videos/v_D-BRqQhL74Q.mp4", "activitynet/videos/v_g1lqpONbOKY.mp4", "activitynet/videos/v_WJ4tCIdAgVo.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A woman plays a black harmonica, briefly pausing before continuing her performance.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The video shows various people playing a flute or similar wind instrument in different settings, with no harmonica visible at any point.\n\n[2] The images focus on a woman in lingerie smoking a cigarette in a dim setting. No musical instruments are seen.\n\n[3] This video shows people playing an accordion. There are no harmonicas visible.\n\n[4] A woman is seated and smoking a cigarette, with no musical instruments apparent.\n\n[5] Displays a woman holding and playing what appears to be a harmonica. She pauses briefly before continuing, matching the query description.\n</content>\n\n<contrast>\n- [5] beats others because it directly shows a woman playing a harmonica with a pause.\n- [1] involves musical performance but does not include a harmonica.\n- [3] displays musical instruments, but they're accordions, not harmonicas.\n- [2] and [4] are irrelevant as they depict smoking without any musical context.\n</contrast>\n\n<summary>\n1st=[5] (Shows a woman playing a harmonica, with a pause as described)\n2nd=[1] (Involves musical performance, albeit with a flute)\n3rd=[3] (Displays instruments but not harmonicas)\n4th=[2] (Unrelated smoking scene)\n5th=[4] (Unrelated smoking scene)\n</summary>\n</think>\n<answer> [5] > [1] > [3] > [2] > [4] </answer>", "role": "assistant" } ]	[ "activitynet/videos/v_gJydcG-d1cs.mp4", "activitynet/videos/v_zuNQFkkyBNo.mp4", "activitynet/videos/v_DPw74KWZzGY.mp4", "activitynet/videos/v_2QO41PbOUJ0.mp4", "activitynet/videos/v_nA9MM4baSdc.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A man attempts to hit a pinata at a group gathering but accidentally strikes someone nearby, which is replayed in slow motion.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] This video depicts a gathering in a backyard setting where a woman is swinging a racket, seemingly not at a pinata, and other activities among the group. There's no clear sequence of someone accidentally hitting another person.\n\n[2] The video involves a person raking leaves in a driveway. There is no pinata or group interaction depicted, and it is unrelated to the query.\n\n[3] Children are trying to hit a pinata in an outdoor setting, but no accidental hits or slow-motion replay of such an event is visible.\n\n[4] Children are gathered around a pinata, but there is no indication of an accidental strike, and no slow-motion aspect is depicted.\n\n[5] A man dressed in a suit is attempting to hit a pinata in front of a group. It appears he accidentally strikes someone nearby, and the scene is played out in slow motion as the person falls over, matching the query description well.\n</content>\n\n<contrast>\n- [5] is more relevant because it clearly shows a man accidentally striking someone nearby with a pinata bat, matching the query, whereas [3] and [4] show pinata activities but lack the accidental strike.\n- [5] includes an individual in a formal setting, adding to the humorous context of an accidental hit being in slow motion.\n- [1] shows a different event, not involving a pinata or any accidental strikes, whereas [5] matches the query directly.\n- [2] is unrelated to pinata or group gatherings, focusing instead on yard work.\n</contrast>\n\n<summary>\n1st=[5] (matches the query of an accidental strike during a pinata activity with slow motion replay)\n2nd=[3] (features pinata activity but lacks accidental strike)\n3rd=[4] (includes pinata but no accident or slow motion)\n4th=[1] (no pinata or relation to the query)\n5th=[2] (completely unrelated to the query)\n</summary>\n</think>\n<answer> [5] > [3] > [4] > [1] > [2] </answer>", "role": "assistant" } ]	[ "activitynet/videos/v_DozJpP2cClI.mp4", "activitynet/videos/v_hQl2eeYX3IY.mp4", "activitynet/videos/v_3HP2E_a9xrU.mp4", "activitynet/videos/v_3fg_CB3rBSk.mp4", "activitynet/videos/v_BJGACTBaU_Y.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A woman in workout attire performs sit-ups while balancing and swinging on a punching bag, demonstrating a creative fitness routine.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] This video shows a woman performing a stepping routine on an aerobic step platform outdoors. She is wearing workout attire and appears to be demonstrating a type of aerobic exercise, but it does not involve a punching bag or creative balancing.\n\n[2] The video depicts two individuals in a martial arts setting, performing acrobatic movements and sparring. While it involves physical activity, it does not show a woman on a punching bag performing sit-ups.\n\n[3] The video shows a woman in workout attire interacting with a punching bag. She is seen engaging with the bag in a dynamic and acrobatic manner, appearing to perform creative fitness movements that could involve balancing and swinging.\n\n[4] This video features a person practicing martial arts moves indoors. The individual is engaged in swinging and striking motions, but there is no punching bag or sit-up exercises involved.\n\n[5] This video shows a woman performing exercises on the floor involving leg raises, but it does not include any interaction with a punching bag or balancing exercises. The exercise is more static compared to the creative dynamic movement described in the query.\n</content>\n\n<contrast>\n- [3] is directly relevant as it shows a woman creatively interacting with a punching bag in a manner that suggests balancing and swinging, satisfying the query.\n- [1] involves aerobic stepping without a punching bag, lacking the described activity of swinging or sit-ups.\n- [5] focuses on floor exercises and doesn't involve a punching bag, whereas [3] shows interaction with one.\n- [4] and [2] involve martial arts and acrobatic maneuvers that are unrelated to the punching bag routine mentioned in the query.\n</contrast>\n\n<summary>\n1st=[3] Clearly shows a woman creatively interacting with a punching bag.\n2nd=[5] Shows fitness activity but without a punching bag.\n3rd=[1] Involves fitness but lacks the dynamic aspect with a punching bag.\n4th=[4] Martial arts focus, lacks relevance to the query.\n5th=[2] Acrobatic sparring, not matching the query's focus.\n</summary>\n</think>\n<answer> [3] > [5] > [1] > [4] > [2] </answer>", "role": "assistant" } ]	[ "activitynet/videos/v_DxNIzbD6uD4.mp4", "activitynet/videos/v_mRyXubjYWEo.mp4", "activitynet/videos/v_t6trrsdhMEc.mp4", "activitynet/videos/v_tilmCzSf1oc.mp4", "activitynet/videos/v_vbWdGj1sfO8.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"An older man enjoys pushing a young girl on a swing, making her smile and laugh as they share a joyful moment together.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The video shows a man pushing a young girl on a swing. The girl seems to be enjoying the swing, smiling, and laughing, but the man does not appear significantly older.\n[2] This video depicts a setting at an outdoor fair where children are riding a camel, guided by adults. It doesn't match the setting or actions described in the query.\n[3] The video involves older adults playing croquet on a lawn, which does not involve a swing or imply the playful interaction described in the query.\n[4] This video appears to be a group activity with a crowd, possibly involving some form of outdoor game, but it does not include swings or a one-on-one interaction between an older man and a young girl.\n[5] The video focuses on a young girl enjoying a swing set independently. There is no older man present to push the swing or interact with the child.\n</content>\n\n<contrast>\n- [1] slightly aligns with the scenario by featuring a man and young girl on a swing, whereas [5] lacks the older man's presence.\n- [5] shows a girl on a swing, matching part of the setting, but [1] includes interaction with another person pushing her.\n- [3] and [4] contain people and outdoor activities but no scene aligns with swinging or direct interaction as in the query.\n- [2] shows children but involves camel rides, lacking the swing interaction described.\n</contrast>\n\n<summary>\n1st=[1] (Includes a man interacting with a girl on a swing with visible enjoyment)\n2nd=[5] (Shows a girl enjoying a swing but lacks an older man's presence)\n3rd=[3] (Outdoor setting with older adults but no swing interaction)\n4th=[4] (Outdoor activity but unrelated to swinging or specific interaction)\n5th=[2] (Children and animals present but setting doesn't match query)\n</summary>\n</think>\n<answer> [1] > [5] > [3] > [4] > [2] </answer>", "role": "assistant" } ]	[ "activitynet/videos/v_dla6qDuAXTw.mp4", "activitynet/videos/v_SHfFrzy_fKM.mp4", "activitynet/videos/v_3YdmY4oPE5s.mp4", "activitynet/videos/v_BJGACTBaU_Y.mp4", "activitynet/videos/v_hDV6wQKvKCg.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A bald man demonstrates a hair cutting technique on a mannequin, showcasing the process and final result for the camera.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The video features a bald man demonstrating a hair cutting technique on a mannequin. The sequence starts with an introduction, followed by cutting and styling processes, and ends with the final hair styling result, which aligns with showcasing the process and final outcome.\n\n[2] A person with short hair is styling a client's hair, but there is no mannequin involved. The setting focuses more on communication and styling tips rather than demonstrative cutting on a mannequin.\n\n[3] This sequence involves a hair treatment and styling process on a human subject, not a mannequin. The focus is on detailing steps for hair styling with tools and products, lacking the specific demonstration on a mannequin required by the query.\n\n[4] The video displays various outdoor activities, mainly skateboarding, with no connection to hair cutting or mannequin demonstrations.\n\n[5] This features a group performing cheerleading stunts. It contains no elements related to hair cutting or mannequin demonstrations.\n</content>\n\n<contrast>\n- [1] beats [2] because it features a bald man demonstrating a hair cutting technique on a mannequin, which perfectly matches the query.\n- [2] beats [3] because it is closer to a hair styling context, while [3] involves styling but not on a mannequin or by a bald man.\n- [3] beats [4] because it is hair-related, whereas [4] has no relation to hair styling or cutting.\n- [4] beats [5] as it features various activities, whereas [5] is entirely unrelated with cheerleading performances.\n- [1] clearly contains all elements of the query including a bald man, mannequin, hair cutting, and showcasing the result.\n</contrast>\n\n<summary>\n1st=[1] (Best match: Bald man, mannequin, hair cutting, and showcasing result)\n2nd=[2] (Hair styling context closer than others)\n3rd=[3] (Involves hair but not with a mannequin or bald man)\n4th=[4] (Unrelated outdoor activities, better than cheerleading which is entirely off-topic)\n5th=[5] (Completely unrelated cheerleading content)\n</summary>\n</think>\n<answer> [1] > [2] > [3] > [4] > [5] </answer>", "role": "assistant" } ]	[ "activitynet/videos/v_JU_o9ZtH-VM.mp4", "activitynet/videos/v_vtYbJD2J8Tw.mp4", "activitynet/videos/v_hPQEkPZmNsU.mp4", "activitynet/videos/v_DXIHbnyfNGA.mp4", "activitynet/videos/v_ujWG6rjlN3s.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A group of men play an outdoor foosball game, celebrating their win with handshakes, hugs, and a high five with a nearby child.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] Displays a group of men playing foosball outdoors, followed by celebratory gestures such as handshakes, hugs, and the involvement of a child with a high five. This matches the query closely.\n[2] Shows two women arm wrestling at a table with a crowd around them. This is unrelated to the query about foosball and celebration.\n[3] Captures a scene involving bulls and people outdoors. There is no relation to foosball or celebrations.\n[4] Features a group playing foosball indoors, displaying medals and certificates. However, it lacks the specific outdoor setting and celebratory gestures mentioned in the query.\n[5] Shows a hockey game with a physical altercation and team gathering, unrelated to foosball or the described celebration.\n\n</content>\n<contrast>\n- [1] is ranked higher than [4] because [1] shows an outdoor foosball game with celebrations including a child, precisely matching the query.\n- [4] is ranked higher than [2] because [4] involves foosball, though not outdoors or with the same celebration dynamics.\n- [2] is ranked higher than [5] because both depict competitive activities, yet [2] is further from the query domain of foosball.\n- [5] is ranked higher than [3] because [5] involves a sport scene, whereas [3] is entirely unrelated to sports or celebrations.\n- [3] is ranked lowest as it depicts a completely unrelated activity involving bulls.\n\n</contrast>\n<summary>\n1st=[1] (matches outdoor foosball game, celebration, child involvement)\n2nd=[4] (features foosball, but indoors without specific celebration details)\n3rd=[2] (competitions but unrelated to foosball)\n4th=[5] (sports scene but unrelated to foosball)\n5th=[3] (completely unrelated to query)\n\n</summary>\n</think>\n<answer> [1] > [4] > [2] > [5] > [3] </answer>", "role": "assistant" } ]	[ "activitynet/videos/v_iC1C5wfb0iU.mp4", "activitynet/videos/v_eil9mWGJB8E.mp4", "activitynet/videos/v_esQnfUcBbPE.mp4", "activitynet/videos/v_mnv-Qz4QDfU.mp4", "activitynet/videos/v_ruHObln9mwI.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A series of humorous and unexpected mishaps occur as various drummers and a marching band perform, resulting in broken drumsticks, falling instruments, and several drummers taking tumbles during their energetic routines.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] This grid shows multiple scenes related to drumming performances. There are several instances where drumsticks break, drummers have surprised reactions, and a marching band is involved. These elements suggest humorous or unexpected mishaps during drumming activities.\n[2] This grid depicts a basketball lay-up drill with instructional text, unrelated to drumming or humorous mishaps.\n[3] This grid focuses on demonstrating footwear functionality on grass, involving close-up shots of shoes, not related to the query.\n[4] The grid shows a marching band performing in a parade. There is no clear evidence of humorous or unexpected events occurring within this sequence.\n[5] This grid showcases a fitness equipment advertisement, unrelated to drumming or mishaps in performance.\n</content>\n\n<contrast>\n- [1] contains humorous drumming mishaps, aligning with the query, whereas [4] only shows a marching band performing without visible mishaps.\n- [2], [3], and [5] are unrelated to drumming or humorous scenarios, focusing instead on a basketball drill, footwear functionality, and fitness advertisement, respectively.\n</contrast>\n\n<summary>\n1st=[1] (contains multiple instances of drumming mishaps and marching band performance, directly related to the query)\n2nd=[4] (shows a marching band, part of the query, but lacks visible humorous mishaps)\n3rd=[2] (unrelated, but coherent content about basketball drills)\n4th=[5] (unrelated fitness equipment advertisement with coherent content)\n5th=[3] (unrelated footwear demonstration with some redundancy)\n</summary>\n</think>\n<answer> [1] > [4] > [2] > [5] > [3] </answer>", "role": "assistant" } ]	[ "activitynet/videos/v_KX-vl2o8U1w.mp4", "activitynet/videos/v_DmTI8y6Anvo.mp4", "activitynet/videos/v_kDurnkUZCXs.mp4", "activitynet/videos/v_jkn6uvmqwh4.mp4", "activitynet/videos/v_pA8QJ2ZoeBM.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A woman at the beach demonstrates mixing sand and water while interacting with the camera and showcasing the surrounding beach activity.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] Shows a diving pool scene with a person diving from a board, no beach interaction or sand and water activity.\n[2] Depicts a person flying a kite on a beach, with no interaction involving sand or water.\n[3] Displays a beach volleyball game, with no focus on mixing sand and water or specific camera interaction.\n[4] Shows a woman at the beach preparing sand, mixing with water, and interacting with the camera, aligning with the query.\n[5] Involves an arm-wrestling competition, unrelated to beach activity or mixing sand and water.\n</content>\n\n<contrast>\n- [4] is the best match as it features a woman mixing sand and water while interacting with the camera, fitting the query perfectly. \n- [2] shows beach activity but lacks the specific sand and water interaction.\n- [3] features beach activity but does not involve mixing sand and water, nor direct camera interaction.\n- [1] and [5] do not relate to beach scenes or mixing sand and water at all.\n</contrast>\n\n<summary>\n1st=[4] (perfect alignment with woman mixing sand and water at the beach with camera interaction)\n2nd=[2] (related to beach activity)\n3rd=[3] (related to beach setting but lacks specific interaction with sand and water)\n4th=[1] (unrelated diving scene)\n5th=[5] (unrelated arm-wrestling event)\n</summary>\n</think>\n<answer> [4] > [2] > [3] > [1] > [5] </answer>", "role": "assistant" } ]	[ "activitynet/videos/v_CIJp-SG3sWE.mp4", "activitynet/videos/v_ABMVo0NDhKo.mp4", "activitynet/videos/v_SLPD1U5iDjY.mp4", "activitynet/videos/v_oZgyW8Yca_M.mp4", "activitynet/videos/v_eil9mWGJB8E.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A group of people enjoy riding bumper cars, maneuvering and overcoming small obstacles until the ride comes to a stop.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] This series portrays individuals actively maneuvering bumper cars in a brightly lit amusement park setting. The sequence shows them engaged in riding, steering, and interacting with obstacles until the cars gradually come to rest.\n\n[2] The video depicts people riding bumper cars in an amusement area. There are several frames of bumper cars in motion with people steering and enjoying the ride. The sequence concludes with people leaving the ride area.\n\n[3] Depicts a snowy urban environment with buses and cars navigating the streets. There is no relevance to bumper car rides.\n\n[4] Displays people, mainly children and teenagers, actively riding and steering bumper cars indoors. The sequence includes interactions, maneuvering, and concludes with the cars coming to a rest.\n\n[5] The video shows a demonstration of an electric car jack and wheel spanner being used on a car. This is unrelated to bumper cars.\n</content>\n\n<contrast>\n- [1] is ranked higher than [2] because it clearly shows bumper cars in continuous motion and gradually stopping, closely matching the query.\n- [2] beats [4] as it includes a clear sequence of bumper cars in motion and a definitive stop, whereas [4] has less emphasis on stopping.\n- [4] is better than [3] due to its relevance to bumper cars, while [3] is unrelated.\n- [3] is ahead of [5] because it involves vehicles, whereas [5] demonstrates a car tool unrelated to any amusement activity.\n</contrast>\n\n<summary>\n1st=[1] (matches bumper cars, motion, obstacles, and stopping)\n2nd=[2] (bumper cars, motion, ends with people leaving)\n3rd=[4] (bumper cars in motion but less emphasis on stopping)\n4th=[3] (urban vehicle scene, not bumper cars)\n5th=[5] (car tool demonstration, unrelated)\n</summary>\n</think>\n<answer> [1] > [2] > [4] > [3] > [5] </answer>", "role": "assistant" } ]	[ "activitynet/videos/v_2fs09c8gFsk.mp4", "activitynet/videos/v_SsW-IeF0bzE.mp4", "activitynet/videos/v_thvpt_lUxTI.mp4", "activitynet/videos/v_E5bRptEXtq8.mp4", "activitynet/videos/v_vWsPUB8dQMU.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"This video provides comprehensive tips and demonstrations on arm wrestling techniques, including proper stance, positioning, and movement, as well as information on how to find local arm wrestling clubs.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] This video is about car wash services, specifically highlighting Eastridge Carwash. It showcases different aspects of car cleaning and advertising their services, containing multiple frames with text-based promotions and images related to car washes.\n[2] The video here demonstrates soccer dribbling techniques. It includes individuals practicing passing and dribbling drills on a soccer field, highlighting various skills related to soccer training.\n[3] The video shows two individuals arm wrestling, providing tips and demonstrations on arm wrestling techniques. The frames include text instructions on positioning and movement, which align with the query about arm wrestling techniques.\n[4] This video depicts an arm wrestling tutorial and training regime. It shows people arm wrestling and lifting weights, focusing on arm strength, techniques, and insights into arm wrestling training. The context matches the query.\n[5] The video involves Tai Chi Combat techniques, demonstrating martial arts moves between two individuals in various scenarios. This content is related to Tai Chi, not arm wrestling or clubs.\n</content>\n\n<contrast>\n- [3] beats [4] because it provides direct visual demonstrations and textual tips focused specifically on arm wrestling techniques, matching the query's requirements for stance, positioning, and movement.\n- [4] beats [5] because it focuses on arm wrestling and includes training footage, meeting more of the query's specifics than Tai Chi demonstrations.\n- [5] beats [2] because martial arts techniques share more aspects with arm wrestling than soccer drills, despite neither matching the query.\n- [2] beats [1] because sports training is more related to the theme of techniques than a car wash advertisement, though both don't meet the query well.\n</contrast>\n\n<summary>\n1st=[3] (Detailed demonstration of arm wrestling techniques with text guidance)\n2nd=[4] (Includes arm wrestling training, relevant to the query)\n3rd=[5] (Martial arts, slight relation to physical techniques)\n4th=[2] (Soccer drills, unrelated to arm wrestling)\n5th=[1] (Car wash ad, completely irrelevant)\n</summary>\n</think>\n<answer> [3] > [4] > [5] > [2] > [1] </answer>", "role": "assistant" } ]	[ "activitynet/videos/v_T-rV3KKFOgg.mp4", "activitynet/videos/v_zGT6EplIOHY.mp4", "activitynet/videos/v_w6kBE7BHHeU.mp4", "activitynet/videos/v_ecVUPpBNpdA.mp4", "activitynet/videos/v_Fp_uU-qMDSc.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A group of people go snowboarding down a snowy hill before loading their gear onto their car and driving away.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] This sequence features a mix of outdoor winter sports and activities, but primarily showcases what appears to be a snowboarding competition. There are frames showing snowboarders, spectators, and a ski lift, indicating a snowboarding event or practice. There is no clear indication of loading gear onto a car or driving away.\n\n[2] This sequence depicts a snowboarding activity down a snowy hill. The frames show individuals snowboarding, culminating with a frame that includes a vehicle, suggesting gear loading or departure. This sequence matches the query well with snowboarding and a vehicle present.\n\n[3] This sequence contains images from an indoor ice hockey setting. Players are on the ice, some appear injured or in a game situation, which does not relate to snowboarding or the outdoor activity described in the query.\n\n[4] The sequence shows a person clearing snow off a car in a snowy environment. While there is a vehicle and snow involved, the activity of snowboarding is absent, and the sequence mainly focuses on clearing snow rather than winter sports.\n\n[5] This sequence features individuals sledding down a snowy hill. There’s no indication of snowboarding or loading gear onto a vehicle. The activity involves sledding, not snowboarding, and lacks a connection to cars and driving away.\n</content>\n\n<contrast>\n- [2] beats [1] because it includes snowboarding on a snowy hill and a vehicle, whereas [1] lacks the vehicle and focuses on a competitive setting.\n- [4] beats [5] because it involves a car and snow, whereas [5] features sledding with no cars.\n- [5] beats [3] as it occurs outdoors in snowy conditions, aligning more with the concept of snowboarding than indoor hockey.\n</contrast>\n\n<summary>\n1st=[2] (complete snowboarding scene with vehicle presence)\n2nd=[1] (snowboarding event without vehicle)\n3rd=[4] (vehicle and snow present, but no snowboarding)\n4th=[5] (snow activity but wrong type)\n5th=[3] (unrelated ice hockey content)\n</summary>\n</think>\n<answer> [2] > [1] > [4] > [5] > [3] </answer>", "role": "assistant" } ]	[ "activitynet/videos/v_FNB1BMEtIvw.mp4", "activitynet/videos/v_vu65aIIJHtU.mp4", "activitynet/videos/v_CEEJ1gzdmAk.mp4", "activitynet/videos/v_xgIxKDm9M38.mp4", "activitynet/videos/v_6TgVNz6PATk.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A man covers a blank wall with wallpaper by rolling down and attaching each frame.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] A person in a room appears to be interacting with an iron and paper on the floor, possibly heating or flattening something. The scene does not involve wallpapering a wall, and the action is on the floor with an unclear purpose.\n\n[2] A man is on a ladder attaching a large printed wallpaper to a blank wall. The sequence shows the wallpaper being methodically aligned and adhered, covering the surface completely. This matches the query of covering a wall with wallpaper.\n\n[3] A child and an adult play hopscotch in a driveway, drawing squares with chalk and interacting with each other. There is no wallpapering activity or wall interaction involved.\n\n[4] A person is applying a coating to a roof with a roller, focusing on treating the surface. This involves spreading material but is not related to wallpapering a wall.\n\n[5] People are shown laying tiles on a floor. The scene involves preparing and placing tiles, not wallpapering a wall.\n</content>\n\n<contrast>\n- [2] is the most relevant because it clearly shows a man attaching wallpaper to a wall, whereas [1] lacks a clear connection to wallpapering and doesn't involve a wall.\n- [3] involves playing hopscotch; unrelated to the query where wallpapering a wall is specified, whereas [2] shows direct application.\n- [4] focuses on coating a roof, which involves spreading but not on a wall or with wallpaper; [2] shows wallpaper against a wall.\n- [5] deals with floor tiling, which is a different surface and process, whereas [2] involves wall-specific wallpaper.\n</contrast>\n\n<summary>\n1st=[2] (A man installing wallpaper on a wall precisely fits the query)\n2nd=[1] (Involves sheets but not on a wall; somewhat related to surface work)\n3rd=[4] (Application of material but not on a wall)\n4th=[5] (Surface covering, but on a floor with tiles)\n5th=[3] (Unrelated activity of hopscotch with no wall or wallpaper)\n</summary>\n</think>\n<answer> [2] > [1] > [4] > [5] > [3] </answer>", "role": "assistant" } ]	[ "activitynet/videos/v_RNrxxPOyHo4.mp4", "activitynet/videos/v_PG0ao4HkF8M.mp4", "activitynet/videos/v_HafSt7EDin0.mp4", "activitynet/videos/v_ir1jpnsK5p8.mp4", "activitynet/videos/v_fVg3StD8LL0.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A man clears snow off his car with a broom while his dog plays nearby, before parking his car on a snowy day.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] This sequence shows a woman using a snowblower to clear snow around a parked car. The process seems to be continuous but no involvement of a broom, dog, or a man is shown. It lacks key elements like a dog or a man with a broom.\n\n[2] A man is seen using a broom to clear snow off a car in a snowy setting. There is a dog visible in one of the frames. The man later moves the car, aligning with the actions described in the query. This candidate captures the main elements and sequence: snow clearing, presence of a dog, and car movement.\n\n[3] A close-up demonstration of sealing procedure with adhesive materials. This is unrelated to snow clearing or any outdoor snowy scene.\n\n[4] A video of roof repair using coatings. It involves outdoor work but not related to snow, cars, or dogs.\n\n[5] A creative demonstration involving ink and brush art projects. This does not relate to the query about snow clearing or a car on a snowy day.\n</content>\n\n<contrast>\n- [2] is the most relevant as it shows a man with a broom, snow, a dog, and a car being moved, whereas [1] lacks a man, broom, and dog.\n- [1] involves snow but features a woman with a snowblower rather than a man with a broom and no dog or parking action, unlike [2].\n- [3] is unrelated as it focuses on sealant procedures indoors.\n- [4] involves outdoor work but is not related to snow or cars.\n- [5] is an art-related video, completely unrelated to the snow clearing query.\n</contrast>\n\n<summary>\n1st=[2] (shows a man clearing snow with a broom, a dog playing, and a car being parked)\n2nd=[1] (involves snow being cleared but lacks a man, dog, or broom)\n3rd=[4] (some outdoor elements but unrelated to snow or cars)\n4th=[3] (completely unrelated indoor sealant work)\n5th=[5] (unrelated creative art project)\n</summary>\n</think>\n<answer> [2] > [1] > [4] > [3] > [5] </answer>", "role": "assistant" } ]	[ "activitynet/videos/v_jml_hnhjltg.mp4", "activitynet/videos/v_xgIxKDm9M38.mp4", "activitynet/videos/v_PzAjZbCEbLU.mp4", "activitynet/videos/v_ir1jpnsK5p8.mp4", "activitynet/videos/v_N_zbcVs419E.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A person uses a paint brush and ink dropper to create artwork on paper, carefully applying and blending ink with a paper towel.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The video shows a person creating artwork using a brush and paint on fabric, forming floral designs in a hoop. It involves applying paint with a brush and does not include an ink dropper or paper towel.\n[2] This video details a cooking process, specifically preparing and baking a cake. Ingredients are mixed, and the result is a finished cake. No art tools, paint, or ink are involved.\n[3] Here, a person carves and decorates pumpkins. The sequence includes drawing and cutting, with tools like a marker and knife but no paintbrush, ink dropper, or paper.\n[4] The video depicts kite making and flying. It features the assembly of the kite using sticks and a garbage bag. There is no painting or drawing involved.\n[5] This sequence shows a person creating artwork on paper using an ink dropper and possibly blending colors. Various animal and plant designs appear, accomplished with artistic tools relevant to the query.\n\n</content>\n\n<contrast>\n- [5] fits because it involves creating artwork on paper using an ink dropper and blending techniques.\n- [1] uses a brush for painting but lacks an ink dropper or paper association.\n- [2] focuses on cake making where no art creation is involved.\n- [3] involves pumpkin carving with drawing but doesn't use paint or blending techniques.\n- [4] features kite creation with no artistic painting elements.\n\n</contrast>\n\n<summary>\n1st=[5] (involves ink dropper, creating artwork on paper, blending)\n2nd=[1] (uses a brush for art but on fabric, not paper)\n3rd=[3] (involves drawing, not relevant tools for painting)\n4th=[2] (unrelated to art, focuses on cooking)\n5th=[4] (unrelated to painting, involves kite building)\n\n</summary>\n</think>\n<answer> [5] > [1] > [3] > [2] > [4] </answer>", "role": "assistant" } ]	[ "activitynet/videos/v_8rypSctbfjU.mp4", "activitynet/videos/v_YS3Q47uz3HI.mp4", "activitynet/videos/v_GldxuTsExZM.mp4", "activitynet/videos/v_AH-3mF6wWhU.mp4", "activitynet/videos/v_N_zbcVs419E.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"The video showcases various flowers, introduces different bee breeds, and provides a lesson on proper tree mulching techniques.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The video depicts a cooking process, showing a person handling ingredients and preparing a dish. It includes chopping, stirring, and presenting pasta. Not relevant to the query about flowers, bees, or tree mulching.\n\n[2] This video shows a person explaining and demonstrating tree mulching techniques. The sequence includes raking mulch and providing instructions near a tree. It directly pertains to the tree mulching part of the query but lacks content about flowers and bees.\n\n[3] The frames present both flowers in pots and a detailed demonstration of tree mulching. It covers two aspects of the query: showcasing flowers and teaching mulching, making it a strong match. However, there is no content related to bee breeds.\n\n[4] The video focuses on painting abstract flowers. It includes scenes of a person painting and detailing various artistic techniques for flower depiction. It relates to the query's mention of flowers, but lacks references to bee breeds or mulching.\n\n[5] The video is about cleaning shoes, showing steps to remove dirt from footwear. It doesn't contain content related to the query's focus on flowers, bees, or tree mulching.\n</content>\n\n<contrast>\n- [3] is ranked higher than [2] because it includes both flower showcasing and tree mulching, whereas [2] only covers mulching.\n- [2] beats [4] because it addresses tree mulching directly, whereas [4] only deals with painting flowers.\n- [4] is higher than [1] and [5] because it at least involves flowers, whereas [1] and [5] have no content related to the query.\n- [1] involves cooking and has no relevance, whereas [5] is about cleaning shoes and also not relevant. However, cooking is typically more engaging than shoe cleaning, providing marginally more potential interest.\n</contrast>\n\n<summary>\n1st=[3] as it showcases flowers and provides mulching techniques.\n2nd=[2] focusing on tree mulching overview and instructions.\n3rd=[4] depicts flower painting, touching on floral themes.\n4th=[1] cooking sequence irrelevant to the query.\n5th=[5] shoe cleaning, unrelated to the query theme.\n</summary>\n</think>\n<answer> [3] > [2] > [4] > [1] > [5] </answer>", "role": "assistant" } ]	[ "activitynet/videos/v_UvPUywSVy1k.mp4", "activitynet/videos/v_B8WIh6PUjE0.mp4", "activitynet/videos/v_YAjqqbSsFTg.mp4", "activitynet/videos/v_fRUoSevi63M.mp4", "activitynet/videos/v_kDurnkUZCXs.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"People dressed in various character costumes energetically compete in a playful lacrosse game on a green field, highlighting fun and friendly rivalry.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The video shows people dressed in various character costumes, energetically participating in a playful lacrosse game on a green field. The playful and friendly rivalry is highlighted as the characters move around excitedly.\n[2] The video consists of a standard lacrosse game, with players dressed in traditional sports uniforms, rather than character costumes. There is no evident playful or costume element.\n[3] The video shows children playing outside at a party setting, focusing on a piñata game rather than lacrosse. It highlights a playful atmosphere but is unrelated to lacrosse or adult costumes.\n[4] The video captures a polo match with players on horses, clearly unrelated to lacrosse or character costumes.\n[5] The video depicts a tug-of-war game on a field, with participants wearing casual clothing, not in costumes, and unrelated to lacrosse.\n</content>\n\n<contrast>\n- [1] is directly relevant because it features people in costumes playing lacrosse, matching both the sport and character costume element, whereas [2] features a traditional lacrosse game without costumes.\n- [2] includes a lacrosse game but lacks any costume elements, hence less relevant than [1].\n- [3] and [4] involve activities completely unrelated to lacrosse or costumes, making them less relevant than [1], [2], or [5].\n- [5] shows a playful field activity but is not lacrosse-related and lacks costume elements, making [1] more relevant.\n</contrast>\n\n<summary>\n1st=[1] (people in costumes playing lacrosse in a playful manner)\n2nd=[2] (lacrosse game without costumes, standard sports setting)\n3rd=[5] (field activity, playful but not lacrosse or costume related)\n4th=[3] (playful party with kids, not lacrosse related)\n5th=[4] (polo game, unrelated to lacrosse or costumes)\n</summary>\n</think>\n<answer> [1] > [2] > [5] > [3] > [4] </answer>", "role": "assistant" } ]	[ "activitynet/videos/v_hK3E6njydQA.mp4", "activitynet/videos/v_Gc1Mk5UyECQ.mp4", "activitynet/videos/v_3fg_CB3rBSk.mp4", "activitynet/videos/v_LgUCZHYBBEc.mp4", "activitynet/videos/v_c_KU8M2YfzA.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A video demonstrates traditional shaving techniques using a boar bristle brush and straight razor, as a barber shaves a man's face and the man shaves his head.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] This sequence involves a stylist working on a mannequin’s hair, showcasing hair cutting techniques. There is no shaving or use of a straight razor.\n\n[2] This sequence shows a woman demonstrating knife sharpening techniques using a sharpening stone. No elements of shaving or use of boar bristle brushes are present.\n\n[3] This sequence depicts a person sharpening a tool with a sharpening stone. No shaving, brushes, or straight razors involved.\n\n[4] This video depicts traditional shaving techniques. It shows a barber using a boar bristle brush to apply shaving cream and a straight razor to shave a man's face, and the man is also shown shaving his head.\n\n[5] Similar to [4], this sequence presents a close-up view of a traditional shave, mostly focusing on shaving a man's face with a straight razor and applying shaving cream using a brush. It does not include head shaving.\n</content>\n\n<contrast>\n- [4] provides a complete demonstration of traditional shaving, including both face and head shaving, whereas [5] focuses mostly on the face.\n- [5] highlights detailed close-ups of the facial shaving process, making it slightly less comprehensive than [4].\n- [1], [2], and [3] are unrelated as they do not involve shaving techniques, boar bristle brushes, or straight razors.\n</contrast>\n\n<summary>\n1st=[4] (comprehensive coverage of face and head shaving using a straight razor and brush)\n2nd=[5] (detailed focus on face shaving with straight razor and brush)\n3rd=[1] (completely unrelated; focus on hair cutting)\n4th=[2] (completely unrelated; focus on knife sharpening)\n5th=[3] (completely unrelated; focus on tool sharpening)\n</summary>\n</think>\n<answer> [4] > [5] > [1] > [2] > [3] </answer>", "role": "assistant" } ]	[ "activitynet/videos/v_JU_o9ZtH-VM.mp4", "activitynet/videos/v_1SQAnbh_lcM.mp4", "activitynet/videos/v_Anvhv9FFv1Q.mp4", "activitynet/videos/v_uvnrLngXHh8.mp4", "activitynet/videos/v_kq7KVmvkYPc.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A woman explains and demonstrates how to section and style another woman's hair using rollers while speaking to the camera.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The video shows a woman demonstrating her blow-drying hair routine, holding hair products, and using a hairbrush and blow dryer on herself. There is no interaction with another person.\n\n[2] A woman is seen sectioning and styling another woman’s hair, employing hair tools and explaining the process. This aligns with the query of demonstrating hair styling using rollers while speaking to the camera.\n\n[3] The video features a woman braiding and pinning her own hair without any interaction with another person or use of rollers.\n\n[4] This video primarily shows a woman discussing various topics and preparing to take a shower. It diverges from the hair styling theme.\n\n[5] The video is about administering mouthwash to another person rather than hair styling. There is no hair-related activity.\n</content>\n\n<contrast>\n- [2] is the most relevant since it shows a woman sectioning and styling another woman's hair while explaining, which matches the query.\n- [1] involves hair styling but focuses on self-demonstration and blow-drying, not using rollers or styling another person's hair.\n- [3] features self-braiding without any instruction on styling another person’s hair.\n- [4] and [5] are unrelated to the hair styling domain, with [4] showing general discussion and [5] focusing on administering mouthwash.\n</contrast>\n\n<summary>\n1st=[2] (matches query with styling another person's hair and explaining)\n2nd=[1] (focuses on self hairstyling, not styling another person)\n3rd=[3] (features self-hair braiding, no interaction)\n4th=[4] (unrelated to hair styling; involves talking and shower preparation)\n5th=[5] (unrelated to hair styling; involves administering mouthwash)\n</summary>\n</think>\n<answer> [2] > [1] > [3] > [4] > [5] </answer>", "role": "assistant" } ]	[ "activitynet/videos/v_RkCSfz8GeNg.mp4", "activitynet/videos/v_vtYbJD2J8Tw.mp4", "activitynet/videos/v_2RpAv3irv1c.mp4", "activitynet/videos/v_MO37MVz4uuA.mp4", "activitynet/videos/v_I-vi5EpjrFI.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A woman walks onto the grass with two small dogs on leashes, and interacts affectionately with them.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] Various scenes of people interacting with dogs outdoors, but no clear sequence of a woman walking two small dogs on leashes. Interaction is affectionate, but not specific to the query.\n\n[2] A person is primarily flying a kite on a large grassy field. Dogs are present but not central, and interaction with them is minimal and not affectionate.\n\n[3] Scenes of horseback riding with dogs visible in the background. No women with leashes or affectionate interaction is depicted.\n\n[4] A man is playing fetch with a dog using a frisbee on grass. The focus is on active play rather than walking or affectionate interaction with small dogs on leashes.\n\n[5] A woman walks two small dogs on leashes around a yard, interacting with them while maintaining control of their leashes. The behavior aligns with the query's description of walking and affectionate interaction.\n</content>\n\n<contrast>\n- [5] shows a woman walking with two dogs on leashes, interacting affectionately, fulfilling the entire query; whereas [1] lacks a clear sequence involving walking two small dogs on leashes.\n- [1] includes affectionate interaction with dogs, but without any walking component, whereas [2] and [3] contain neither.\n- [4] focuses on play rather than walking and lacks the specified affectionate interaction seen in [5].\n</contrast>\n\n<summary>\n1st=[5] (captures the walking and interaction with two small dogs on leashes)\n2nd=[1] (affectionate dog interaction but missing walking aspect)\n3rd=[4] (features interaction with a dog, lacks leashes and the specified interaction type)\n4th=[2] (minimal relevance to query with no dog-centered interaction)\n5th=[3] (focuses on horseback riding, no relevant interaction)\n</summary>\n</think>\n<answer> [5] > [1] > [4] > [2] > [3] </answer>", "role": "assistant" } ]	[ "activitynet/videos/v_cUEGoRjraxM.mp4", "activitynet/videos/v__O6ONcIPR_s.mp4", "activitynet/videos/v_AzUK4Nwd-sY.mp4", "activitynet/videos/v_tCkHrK6mRME.mp4", "activitynet/videos/v_rLH89pN9I84.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"Two men work together on a roof, nailing long shingles along the perimeter.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] Two men are working on staining a fence. They are shown using equipment, with a ladder visible against the fence. The actions involve maintenance or improvement but not roofing or shingling.\n\n[2] A man is chopping wood in a backyard. The sequence shows him preparing and splitting logs, not related to roofing or working with shingles.\n\n[3] This sequence shows two men on a roof laying out and securing long shingles along the perimeter. It matches closely with the query description of men working on a roof with shingles.\n\n[4] Two men are on a roof, seemingly involved in a repair or dismantling process. They are handling roofing materials but do not appear to be nailing long shingles along the perimeter.\n\n[5] The sequence involves roof repair, focusing on patching and repairing valleys on roof pitches. The main action is not laying long shingles, but repair work.\n</content>\n\n<contrast>\n- [3] is highly relevant because it shows two men actively laying shingles on a roof, matching the query directly.\n- [4] involves roof work and shows two men, but the action appears to be repair rather than laying shingles, which is less aligned with the query.\n- [5] is related to roofing but focuses on repair tasks, not laying shingles, making it less relevant than [3] and [4].\n- [1] and [2] show actions unrelated to roofing or shingling, with [1] focusing on fence staining and [2] on wood chopping.\n</contrast>\n\n<summary>\n1st=[3] (directly matches the query with roof shingling activity)\n2nd=[4] (shows roof work, but likely repairing/removing, not laying shingles)\n3rd=[5] (roof repair, less relevant to laying shingles)\n4th=[1] (unrelated action, fence staining)\n5th=[2] (unrelated action, wood chopping)\n</summary>\n</think>\n<answer> [3] > [4] > [5] > [1] > [2] </answer>", "role": "assistant" } ]	[ "activitynet/videos/v_ggGw9EGSMs8.mp4", "activitynet/videos/v_b2uAgV5n3FU.mp4", "activitynet/videos/v_r7E7cVoJ0JE.mp4", "activitynet/videos/v_Cj3W6CW3Zbc.mp4", "activitynet/videos/v_9mpPAwSmUmk.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"The video follows a rapper interacting with people in an urban setting, getting a tattoo, and showcasing artwork, graffiti, and tattoos while highlighting his finished tattoo and album cover.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The video is about arm wrestling with scenes from a gym. There's no relation to a rapper, tattoos, graffiti, or album covers.\n\n[2] The video demonstrates an electric car jack and wheel spanner with various steps shown in a parking lot. It doesn't relate to the query regarding rappers or tattoos.\n\n[3] The video involves a group of people working in a garden or park setting, shoveling mulch. There's no indication of tattoos, rappers, or album covers.\n\n[4] The video depicts a person dancing in a dimly lit area with onlookers, possibly in a bar. It doesn't show interactions or tattoos related to the query.\n\n[5] This video involves a person interacting with others in an urban setting, getting a tattoo, and highlighting artwork. This matches the query well in terms of tattoos and urban interaction.\n</content>\n\n<contrast>\n- [5] beats others because it aligns with the query showing a rapper interacting with people, getting a tattoo, and showcasing artwork.\n- [4] ranks lower because it lacks elements like tattoos or artwork related to a rapper.\n- [3] is about gardening, whereas [5] involves urban settings and tattoos, making it irrelevant.\n- [2] shows a car tool demonstration, not related to rappers or tattoos like [5].\n- [1] focuses on arm wrestling training, which doesn't match the rapper and tattoo theme of [5].\n</contrast>\n\n<summary>\n1st=[5] (Relevant scenes with tattoos, artwork, and urban setting)\n2nd=[4] (Social interaction but no rapper or tattoo relevance)\n3rd=[3] (Unrelated gardening activity)\n4th=[2] (Car tool demonstration, irrelevant to the query)\n5th=[1] (Gym and arm wrestling, no connection to rapper theme)\n</summary>\n</think>\n<answer> [5] > [4] > [3] > [2] > [1] </answer>", "role": "assistant" } ]	[ "activitynet/videos/v_ecVUPpBNpdA.mp4", "activitynet/videos/v_vWsPUB8dQMU.mp4", "activitynet/videos/v_3C00m8lkJi0.mp4", "activitynet/videos/v_Xj--zKhm6k4.mp4", "activitynet/videos/v_-8FS4LbXAgE.mp4" ]

No dataset card yet

Downloads last month: 26

Size of downloaded dataset files:

168 kB

Size of the auto-converted Parquet files:

58.1 kB

Number of rows:

Collection including happy8825/anet_sampled_sft_valid

Video-Retrieval-Dataset

Collection

4 items • Updated 9 days ago