user
string | videos
list | gt
int64 | predict
string | pred_order
list | top1_correct
bool | gt_pos
int64 | latency_ms
float64 | index
int64 | request_messages
list |
|---|---|---|---|---|---|---|---|---|---|
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "Several blocks fall, forming a black and white picture of two girls. The girls are shown in a gymnasium. One does several back flips, then a group begins throwing girls into the air, practicing flips. Several clips are shown of people doing gymnastics and falling down."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_GSFyEkGCUVo.mp4",
"activitynet/videos/v_dSF2i1OQtMc.mp4",
"activitynet/videos/v_m3yLm_dJU94.mp4",
"activitynet/videos/v_iA2Q4t-o58w.mp4",
"activitynet/videos/v_SyOdA4ZKEtQ.mp4"
] | 4
|
<answer>4</answer>
|
[
4
] | true
| 1
| 18,636.305621
| 4,800
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"Several blocks fall, forming a black and white picture of two girls. The girls are shown in a gymnasium. One does several back flips, then a group begins throwing girls into the air, practicing flips. Several clips are shown of people doing gymnastics and falling down.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_GSFyEkGCUVo.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_dSF2i1OQtMc.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_m3yLm_dJU94.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_iA2Q4t-o58w.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_SyOdA4ZKEtQ.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A close up of tin foil is shown leading a woman taking a large hit out of a hookah hose. She continues smoking out of the hookah as the camera captures her movements and she sits down in the end to smoke more."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_v3t4Z5cEgZM.mp4",
"activitynet/videos/v_gWz4P3Jnis8.mp4",
"activitynet/videos/v_DUb48prwNZk.mp4",
"activitynet/videos/v_uIOIcv5MhuA.mkv",
"activitynet/videos/v_Tv8r1w-rLME.mkv"
] | 3
|
<answer>3</answer>
|
[
3
] | true
| 1
| 13,500.959483
| 4,802
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A close up of tin foil is shown leading a woman taking a large hit out of a hookah hose. She continues smoking out of the hookah as the camera captures her movements and she sits down in the end to smoke more.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_v3t4Z5cEgZM.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_gWz4P3Jnis8.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_DUb48prwNZk.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_uIOIcv5MhuA.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_Tv8r1w-rLME.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "The arm of a backhoe in shovel configuration peels the tiles off a roof. A man walks across the frame. The man pitchforks debris into the backhoe shovel."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_UWTpfygMUQw.mp4",
"activitynet/videos/v_stVRtmxHVaE.mp4",
"activitynet/videos/v_aPzHheM0Egw.mp4",
"activitynet/videos/v_4cd0sNdLmT4.mp4",
"activitynet/videos/v_JkZZvDHTty4.mp4"
] | 5
|
<answer>4</answer>
|
[
4
] | false
| 0
| 13,787.029438
| 4,801
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"The arm of a backhoe in shovel configuration peels the tiles off a roof. A man walks across the frame. The man pitchforks debris into the backhoe shovel.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_UWTpfygMUQw.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_stVRtmxHVaE.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_aPzHheM0Egw.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_4cd0sNdLmT4.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_JkZZvDHTty4.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A large group of people are seen standing in a circle surrounding two people in the middle. More people come into frame and the people then perform a dance routine together. The people continuing dancing and clapping around one another and end by holding a pose in the end and bowing to the audience."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_BtrGC6PUPJk.mp4",
"activitynet/videos/v_tm_CL7A0W4M.mp4",
"activitynet/videos/v_dnzcNZBtUG4.mp4",
"activitynet/videos/v_vKCxWIzJTm0.mp4",
"activitynet/videos/v_JhihdPxI_Xc.mp4"
] | 4
|
<answer>2</answer>
|
[
2
] | false
| 0
| 17,618.013559
| 4,803
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A large group of people are seen standing in a circle surrounding two people in the middle. More people come into frame and the people then perform a dance routine together. The people continuing dancing and clapping around one another and end by holding a pose in the end and bowing to the audience.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_BtrGC6PUPJk.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_tm_CL7A0W4M.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_dnzcNZBtUG4.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_vKCxWIzJTm0.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_JhihdPxI_Xc.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "We see an opening title screen. A little girl with wet hair is playing shuffle board on a blue court outdoors. We see a little girl spinning around and around and we see a girl in a yellow dress playing shuffleboard. The girl in pink and the one in yellow walk to a man in a blue shirt and get candy and hugs and kisses. We then see the ending screen."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_fD9JNH5FWCk.mp4",
"activitynet/videos/v_PoamN_DEInI.mp4",
"activitynet/videos/v_YozbZM_nA0c.mp4",
"activitynet/videos/v_B_9S_qzlD38.mp4",
"activitynet/videos/v_t_Creyg6ANs.mp4"
] | 5
|
<answer>5</answer>
|
[
5
] | true
| 1
| 19,486.094253
| 4,805
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"We see an opening title screen. A little girl with wet hair is playing shuffle board on a blue court outdoors. We see a little girl spinning around and around and we see a girl in a yellow dress playing shuffleboard. The girl in pink and the one in yellow walk to a man in a blue shirt and get candy and hugs and kisses. We then see the ending screen.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_fD9JNH5FWCk.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_PoamN_DEInI.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_YozbZM_nA0c.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_B_9S_qzlD38.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_t_Creyg6ANs.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A close up of a floor is seen followed by a person walking down the street. Two people are then seen break dancing with one another and moving around in different locations. The two lift weights and move around areas while moving around areas."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_oUdEoaKDHpA.mp4",
"activitynet/videos/v_UN0bAa_ko4I.mp4",
"activitynet/videos/v_CoP3xaSZt7A.mkv",
"activitynet/videos/v_0ZXc2fEDgg8.mp4",
"activitynet/videos/v_vKCxWIzJTm0.mp4"
] | 1
|
<answer>5</answer>
|
[
5
] | false
| 0
| 27,136.332618
| 4,804
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A close up of a floor is seen followed by a person walking down the street. Two people are then seen break dancing with one another and moving around in different locations. The two lift weights and move around areas while moving around areas.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_oUdEoaKDHpA.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_UN0bAa_ko4I.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_CoP3xaSZt7A.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_0ZXc2fEDgg8.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_vKCxWIzJTm0.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "Four women are standing on a colorful stage. They begin to dance in unison. They gyrate and move side to side, belly dancing."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_bwRsZtPzipc.mkv",
"activitynet/videos/v_sxQ9H3c5bRM.mp4",
"activitynet/videos/v_KhAtzEJxz9M.mp4",
"activitynet/videos/v_1cLxW-FhgpA.mp4",
"activitynet/videos/v_ksvK_P-Eas4.mp4"
] | 5
|
<answer>4</answer>
|
[
4
] | false
| 0
| 13,631.750963
| 4,806
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"Four women are standing on a colorful stage. They begin to dance in unison. They gyrate and move side to side, belly dancing.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_bwRsZtPzipc.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_sxQ9H3c5bRM.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_KhAtzEJxz9M.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_1cLxW-FhgpA.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_ksvK_P-Eas4.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A woman in a brown dress is talking. A woman in a black shirt is standing next to her. There is a table with products on it in front of them. The woman in the black shirt is mixing a paste together in a bowl. She rubs that paste onto a dirty shoe. The other woman holds up a pair of dirty sandals. The women give each other a kiss on the cheek. The woman in the brown dress continues talking. She then rubs the paste onto the shoe again."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_uyGxlvak-Bg.mkv",
"activitynet/videos/v_NDyc4PZE954.mp4",
"activitynet/videos/v_hDPLy21Yyuk.mp4",
"activitynet/videos/v_U1nvAxorOPQ.mp4",
"activitynet/videos/v_vJiOYQE9tts.mp4"
] | 1
|
<answer>1</answer>
|
[
1
] | true
| 1
| 17,152.742132
| 4,807
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A woman in a brown dress is talking. A woman in a black shirt is standing next to her. There is a table with products on it in front of them. The woman in the black shirt is mixing a paste together in a bowl. She rubs that paste onto a dirty shoe. The other woman holds up a pair of dirty sandals. The women give each other a kiss on the cheek. The woman in the brown dress continues talking. She then rubs the paste onto the shoe again.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_uyGxlvak-Bg.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_NDyc4PZE954.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_hDPLy21Yyuk.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_U1nvAxorOPQ.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_vJiOYQE9tts.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A counter top with lemons and ingredients in small bowls is shown, with several individuals standing by the counter. One of the individuals extends a hand and manipulates the ingredients. The individual uses a knife to cut the lemons. The individual squeezes the lemons' juice into a mug using a tool. The individual puts some of the ingredients in small bowls into the mug. The individual rearranges the items on the counter top. The individual mixes some water into the mug. The individual squeezes lemons into the mug. The individual mixes the contents of the mug. Eight containers of liquid are shown on the counter top."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_ADUmfTuiDH8.mp4",
"activitynet/videos/v_zgnBeiEB5pE.mp4",
"activitynet/videos/v_JMpwIWxoB5s.mp4",
"activitynet/videos/v_C0MIMsY6okw.mp4",
"activitynet/videos/v_9bERRZ2eTbo.mp4"
] | 4
|
<answer>3</answer>
|
[
3
] | false
| 0
| 21,457.710934
| 4,808
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A counter top with lemons and ingredients in small bowls is shown, with several individuals standing by the counter. One of the individuals extends a hand and manipulates the ingredients. The individual uses a knife to cut the lemons. The individual squeezes the lemons' juice into a mug using a tool. The individual puts some of the ingredients in small bowls into the mug. The individual rearranges the items on the counter top. The individual mixes some water into the mug. The individual squeezes lemons into the mug. The individual mixes the contents of the mug. Eight containers of liquid are shown on the counter top.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_ADUmfTuiDH8.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_zgnBeiEB5pE.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_JMpwIWxoB5s.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_C0MIMsY6okw.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_9bERRZ2eTbo.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A man is on a raft in a river. He is kayaking through the rough waters. He goes around a bend in the river, recording the view from behind."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_HWuRcUpcsHY.mp4",
"activitynet/videos/v_k_xDTGiDp9A.mp4",
"activitynet/videos/v_h67ctuwV-Nw.mp4",
"activitynet/videos/v_hchuVbHYK_k.mp4",
"activitynet/videos/v_RpyIg_j4I3E.mkv"
] | 4
|
<answer>1</answer>
|
[
1
] | false
| 0
| 26,316.983604
| 4,809
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A man is on a raft in a river. He is kayaking through the rough waters. He goes around a bend in the river, recording the view from behind.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_HWuRcUpcsHY.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_k_xDTGiDp9A.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_h67ctuwV-Nw.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_hchuVbHYK_k.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_RpyIg_j4I3E.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A person is seen putting dough over an outline in fast motion and then rolling out more dough. The person pokes holds in the dough and then paints all over the sides of the cake. She is then seen again putting objects of dough on the cake, rolling out more, and then putting the pieces on the side. In the end she shows off the finished cake."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_q8c_0JTe5r8.mp4",
"activitynet/videos/v_ss6XN-JP_x8.mp4",
"activitynet/videos/v_kxfOrs5ZWkw.mp4",
"activitynet/videos/v_Zr1xfVeUGeo.mp4",
"activitynet/videos/v_J2gJYNO2qh8.mp4"
] | 3
|
<answer>1</answer>
|
[
1
] | false
| 0
| 25,538.90038
| 4,810
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A person is seen putting dough over an outline in fast motion and then rolling out more dough. The person pokes holds in the dough and then paints all over the sides of the cake. She is then seen again putting objects of dough on the cake, rolling out more, and then putting the pieces on the side. In the end she shows off the finished cake.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_q8c_0JTe5r8.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_ss6XN-JP_x8.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_kxfOrs5ZWkw.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_Zr1xfVeUGeo.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_J2gJYNO2qh8.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A man uses a putty knife you scrape a blemish on a surface. The man sues a towel and spray bottle to clean of the top of a cabinet. Paint is applied to a brush and the man puts a first coat onto the bare cabinet. The man stirs a can of paint then applies a second coat onto the cabinet. The man pours out a clear liquid and washes off his brush with it."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_4zYY4abpCgI.mp4",
"activitynet/videos/v_5vDPgcyRtOU.mp4",
"activitynet/videos/v_rBJBnf4F9sA.mkv",
"activitynet/videos/v_dWJIJM3qmyQ.mkv",
"activitynet/videos/v_oFh_AGspaEQ.mp4"
] | 4
|
<answer>4</answer>
|
[
4
] | true
| 1
| 13,643.820595
| 4,811
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A man uses a putty knife you scrape a blemish on a surface. The man sues a towel and spray bottle to clean of the top of a cabinet. Paint is applied to a brush and the man puts a first coat onto the bare cabinet. The man stirs a can of paint then applies a second coat onto the cabinet. The man pours out a clear liquid and washes off his brush with it.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_4zYY4abpCgI.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_5vDPgcyRtOU.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_rBJBnf4F9sA.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_dWJIJM3qmyQ.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_oFh_AGspaEQ.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "The athlete in orange top ran towards the mat, then did a high jump and he was able to go over the pole and landed on the mat. The man in gray shirt on the side is holding a long pole, while a bald man is raising the flags. The athlete picked up the stick and did another high jump."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_ix40OdQd7iE.mp4",
"activitynet/videos/v_XZZRyOhxQBE.mp4",
"activitynet/videos/v_Usowsx0PDbA.mp4",
"activitynet/videos/v_5rw6n16ILgY.mp4",
"activitynet/videos/v_FsS8cQbfKTQ.mp4"
] | 2
|
<answer>3</answer>
|
[
3
] | false
| 0
| 21,355.924547
| 4,812
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"The athlete in orange top ran towards the mat, then did a high jump and he was able to go over the pole and landed on the mat. The man in gray shirt on the side is holding a long pole, while a bald man is raising the flags. The athlete picked up the stick and did another high jump.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_ix40OdQd7iE.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_XZZRyOhxQBE.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_Usowsx0PDbA.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_5rw6n16ILgY.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_FsS8cQbfKTQ.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A person is seen bending over a floor placing tiles down over the plaster. The person continues laying tiles down and pushing down on the floor to make sure it's sturdy."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_WQmJrfjOF7o.mp4",
"activitynet/videos/v_mgNfayAiTQc.mp4",
"activitynet/videos/v_Sh8r9g_lp7U.mp4",
"activitynet/videos/v_dcMFJ-8Eo7g.mp4",
"activitynet/videos/v_djpr7UMlnSw.mp4"
] | 5
|
<answer>2</answer>
|
[
2
] | false
| 0
| 16,120.443858
| 4,813
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A person is seen bending over a floor placing tiles down over the plaster. The person continues laying tiles down and pushing down on the floor to make sure it's sturdy.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_WQmJrfjOF7o.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_mgNfayAiTQc.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_Sh8r9g_lp7U.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_dcMFJ-8Eo7g.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_djpr7UMlnSw.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "little kid is playing in a playground in a big park. woman is sitting with two kids in the lawn."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_Fmr6mPyvE-g.mp4",
"activitynet/videos/v_tAEGMVLn0wk.mp4",
"activitynet/videos/v_yeWCfvmeUvM.mp4",
"activitynet/videos/v_lLHAzwAs_9I.mp4",
"activitynet/videos/v_iwMXYbYyJy4.mp4"
] | 1
|
<answer>2</answer>
|
[
2
] | false
| 0
| 21,669.978751
| 4,814
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"little kid is playing in a playground in a big park. woman is sitting with two kids in the lawn.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_Fmr6mPyvE-g.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_tAEGMVLn0wk.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_yeWCfvmeUvM.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_lLHAzwAs_9I.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_iwMXYbYyJy4.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A close up of a person's hands are seen followed by a man walking into frame. The man is seen holding an instrument and begins to play. The man continues playing the bagpipes and ends by turning off the camera."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_RfchfRzuV8I.mp4",
"activitynet/videos/v_UR1e1MIRvvc.mp4",
"activitynet/videos/v_lDJpGI4BZ8k.mkv",
"activitynet/videos/v_p1gH8y8X0kA.mp4",
"activitynet/videos/v_BP9MfTepAv4.mp4"
] | 1
|
<answer>4</answer>
|
[
4
] | false
| 0
| 13,460.832724
| 4,815
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A close up of a person's hands are seen followed by a man walking into frame. The man is seen holding an instrument and begins to play. The man continues playing the bagpipes and ends by turning off the camera.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_RfchfRzuV8I.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_UR1e1MIRvvc.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_lDJpGI4BZ8k.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_p1gH8y8X0kA.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_BP9MfTepAv4.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A person in a gym describes how to use a machine on your knees. The man pulls the bar and shows the proper technique."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v__1vYKA7mNLI.mp4",
"activitynet/videos/v_cBAlXvu38dg.mp4",
"activitynet/videos/v_phg81-nhqH4.mp4",
"activitynet/videos/v_qjacthwabek.mp4",
"activitynet/videos/v_DV1ITGBfo5w.mp4"
] | 1
|
<answer>1</answer>
|
[
1
] | true
| 1
| 10,036.412616
| 4,816
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A person in a gym describes how to use a machine on your knees. The man pulls the bar and shows the proper technique.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v__1vYKA7mNLI.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_cBAlXvu38dg.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_phg81-nhqH4.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_qjacthwabek.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_DV1ITGBfo5w.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A female gymnast mounts a beam, doing a hand stand. She jumps, then does the splits. She flips several times before dismounting."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_PZ1FVhgTRWU.mp4",
"activitynet/videos/v_oS4w7-0aH8Q.mp4",
"activitynet/videos/v_Y1UwPTU61uk.mp4",
"activitynet/videos/v_c-8GvZKndyQ.mp4",
"activitynet/videos/v_JmtcnoHa66U.mkv"
] | 5
|
<answer>1</answer>
|
[
1
] | false
| 0
| 13,241.254995
| 4,818
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A female gymnast mounts a beam, doing a hand stand. She jumps, then does the splits. She flips several times before dismounting.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_PZ1FVhgTRWU.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_oS4w7-0aH8Q.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_Y1UwPTU61uk.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_c-8GvZKndyQ.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_JmtcnoHa66U.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A man is standing on a kite surf board. He holds onto the sail of the board. Another man in a black wet suit stands in the water. A boat passed by behind him."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_99xnJSBRzkE.mp4",
"activitynet/videos/v_s5y4xXcphcc.mp4",
"activitynet/videos/v_TxgvL4ZJZbo.mp4",
"activitynet/videos/v_jM3Buw2Kidk.mp4",
"activitynet/videos/v_CqscMsSNiNY.mp4"
] | 5
|
<answer>4</answer>
|
[
4
] | false
| 0
| 19,827.45847
| 4,819
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A man is standing on a kite surf board. He holds onto the sail of the board. Another man in a black wet suit stands in the water. A boat passed by behind him.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_99xnJSBRzkE.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_s5y4xXcphcc.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_TxgvL4ZJZbo.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_jM3Buw2Kidk.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_CqscMsSNiNY.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "People sail on inflatable boats down the street in a river. A group of people take a rest having some drinks. After, people sail down the river drinking from cans. A man swing holding a rope attached to a tree. Again, people take a rest while drinking, after they continue sailing on the boats."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_wnnoaLzYqVQ.mp4",
"activitynet/videos/v_yJSQmNSFlNI.mkv",
"activitynet/videos/v_ejIEsnkvLWY.mkv",
"activitynet/videos/v_FX4inHmWQtE.mp4",
"activitynet/videos/v_BKdKbFPerGo.mp4"
] | 2
|
<answer>4</answer>
|
[
4
] | false
| 0
| 29,481.019621
| 4,817
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"People sail on inflatable boats down the street in a river. A group of people take a rest having some drinks. After, people sail down the river drinking from cans. A man swing holding a rope attached to a tree. Again, people take a rest while drinking, after they continue sailing on the boats.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_wnnoaLzYqVQ.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_yJSQmNSFlNI.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_ejIEsnkvLWY.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_FX4inHmWQtE.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_BKdKbFPerGo.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A yellow cheese cake on the plate is shown. The yellow cream is being mixed by a mixer, an egg yolk is added and then mixed it, another egg yolk, then vanilla, then mixed, the flour is sifted and added gradually in the cream and mixed by spatula, then milk is added, and then the sifter flour is also added and then the mixtures are mixed together with a spatula. The egg white is beaten and sugar is added, and then added to the yellow mixture, then it was placed in a pan and baked it."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_gvr1dpCpvhw.mp4",
"activitynet/videos/v_krNVpENNPCM.mp4",
"activitynet/videos/v_yyCsQ7QzAJ8.mp4",
"activitynet/videos/v_nKa1e_CpvoY.mp4",
"activitynet/videos/v_gLfIPN_WM48.mp4"
] | 2
|
<answer>4</answer>
|
[
4
] | false
| 0
| 18,754.632325
| 4,820
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A yellow cheese cake on the plate is shown. The yellow cream is being mixed by a mixer, an egg yolk is added and then mixed it, another egg yolk, then vanilla, then mixed, the flour is sifted and added gradually in the cream and mixed by spatula, then milk is added, and then the sifter flour is also added and then the mixtures are mixed together with a spatula. The egg white is beaten and sugar is added, and then added to the yellow mixture, then it was placed in a pan and baked it.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_gvr1dpCpvhw.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_krNVpENNPCM.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_yyCsQ7QzAJ8.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_nKa1e_CpvoY.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_gLfIPN_WM48.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A little girl is walking in a public playground. She walks beside the sidewalk, looking at the equipment. She climbs up a ladder while her mother attempts to help her, but changes her mind and walks away indecisively. She returns to the ladder, then sits down."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_ccfffP3pXrc.mp4",
"activitynet/videos/v_uKzelWWaYB4.mp4",
"activitynet/videos/v_JJ811udnROI.mp4",
"activitynet/videos/v_tAEGMVLn0wk.mp4",
"activitynet/videos/v_CCRPXH8ui-s.mp4"
] | 4
|
<answer>1</answer>
|
[
1
] | false
| 0
| 19,773.577004
| 4,821
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A little girl is walking in a public playground. She walks beside the sidewalk, looking at the equipment. She climbs up a ladder while her mother attempts to help her, but changes her mind and walks away indecisively. She returns to the ladder, then sits down.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_ccfffP3pXrc.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_uKzelWWaYB4.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_JJ811udnROI.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_tAEGMVLn0wk.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_CCRPXH8ui-s.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A man is seen bending down in the middle of a forest. The man is holding an ax in his hands and begins swinging down. Finally he swings an ax onto a log and walks away."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_MT852hP9wVk.mp4",
"activitynet/videos/v_m6yPz9fHJnY.mp4",
"activitynet/videos/v_cfaBPxE-A5k.mp4",
"activitynet/videos/v_LZC9MLWo9bE.mp4",
"activitynet/videos/v_2-S2fehRKVc.mp4"
] | 5
|
<answer>4</answer>
|
[
4
] | false
| 0
| 20,097.174098
| 4,822
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A man is seen bending down in the middle of a forest. The man is holding an ax in his hands and begins swinging down. Finally he swings an ax onto a log and walks away.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_MT852hP9wVk.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_m6yPz9fHJnY.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_cfaBPxE-A5k.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_LZC9MLWo9bE.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_2-S2fehRKVc.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "Two men are seen kneeling down in front of one another and leads into a wrestling match between the two. One man throws the other down and leads into several more shots of two men wrestling."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_5zT1GWfmVLU.mp4",
"activitynet/videos/v_3KmMvfdidvQ.mp4",
"activitynet/videos/v_0zjA3KPnLK8.mp4",
"activitynet/videos/v_i4SNM6xSLI8.mp4",
"activitynet/videos/v_r8hXEpP7HH0.mp4"
] | 3
|
<answer>2</answer>
|
[
2
] | false
| 0
| 19,502.156218
| 4,823
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"Two men are seen kneeling down in front of one another and leads into a wrestling match between the two. One man throws the other down and leads into several more shots of two men wrestling.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_5zT1GWfmVLU.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_3KmMvfdidvQ.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_0zjA3KPnLK8.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_i4SNM6xSLI8.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_r8hXEpP7HH0.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "The person is shown bathing the dog with the detachable shower head. The camera then zooms in on the dog showing its tail, back, and head. Next the dog gets a white towel wrapped around the bottom half of its body to get dried off. The dog's face is then being held by the person who bathed him, and the camera zooms in again to show the dog sticking his tongue out."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_zyylgHTPUS8.mp4",
"activitynet/videos/v_RN2QwhcAsUY.mp4",
"activitynet/videos/v_6EWzgWd72Cs.mp4",
"activitynet/videos/v_6-beYw2R10s.mp4",
"activitynet/videos/v_Xq9ueKle4fY.mp4"
] | 2
|
<answer>2</answer>
|
[
2
] | true
| 1
| 12,289.810565
| 4,824
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"The person is shown bathing the dog with the detachable shower head. The camera then zooms in on the dog showing its tail, back, and head. Next the dog gets a white towel wrapped around the bottom half of its body to get dried off. The dog's face is then being held by the person who bathed him, and the camera zooms in again to show the dog sticking his tongue out.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_zyylgHTPUS8.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_RN2QwhcAsUY.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_6EWzgWd72Cs.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_6-beYw2R10s.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_Xq9ueKle4fY.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "Two reporters talk in a TV set. Girls talk in a beach, then girls spray sunscreen on the back of women, also a woman wearing a white coat talks and shows products. A man sprays sunscreen on his back, and then he sprays sunscreen to the back of a woman. After, a man sprays sunscreen on his arm and back, other people spray sunscreen on their legs. The reporters in the TV set continue talking."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_yl3bjdUZrmM.mp4",
"activitynet/videos/v_aNE5ZWD5E34.mp4",
"activitynet/videos/v_7S15OsGinjw.mp4",
"activitynet/videos/v_No5ZwqHdEQU.mp4",
"activitynet/videos/v_igrjxhf0XyY.mp4"
] | 2
|
<answer>2</answer>
|
[
2
] | true
| 1
| 13,267.333464
| 4,825
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"Two reporters talk in a TV set. Girls talk in a beach, then girls spray sunscreen on the back of women, also a woman wearing a white coat talks and shows products. A man sprays sunscreen on his back, and then he sprays sunscreen to the back of a woman. After, a man sprays sunscreen on his arm and back, other people spray sunscreen on their legs. The reporters in the TV set continue talking.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_yl3bjdUZrmM.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_aNE5ZWD5E34.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_7S15OsGinjw.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_No5ZwqHdEQU.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_igrjxhf0XyY.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A close up of a contact case is shown and leads into someone washing their hands and putting the contact lens into her eyes. She shows off her bottles and zooms in on the sink and counter top."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_0pcrpO0Gd8M.mp4",
"activitynet/videos/v_gA7GpvB10UY.mkv",
"activitynet/videos/v_HsGz6S2MBU4.mp4",
"activitynet/videos/v_Uj1QtIM8500.mp4",
"activitynet/videos/v_ihdkXBpzKbE.mp4"
] | 2
|
<answer>2</answer>
|
[
2
] | true
| 1
| 11,409.387491
| 4,827
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A close up of a contact case is shown and leads into someone washing their hands and putting the contact lens into her eyes. She shows off her bottles and zooms in on the sink and counter top.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_0pcrpO0Gd8M.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_gA7GpvB10UY.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_HsGz6S2MBU4.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_Uj1QtIM8500.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_ihdkXBpzKbE.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A woman is seen holding onto her hair and looking off into the distance. She then uses a brush to push along her hair. She continues brushing her hair while using her hands."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_-mX18jJkPDk.mp4",
"activitynet/videos/v_7O9kkDxEvaY.mp4",
"activitynet/videos/v_MaJlWFemO68.mp4",
"activitynet/videos/v_hiVs1hNyPpw.mp4",
"activitynet/videos/v_sdR443ncw-I.mp4"
] | 3
|
<answer>2</answer>
|
[
2
] | false
| 0
| 9,628.8373
| 4,828
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A woman is seen holding onto her hair and looking off into the distance. She then uses a brush to push along her hair. She continues brushing her hair while using her hands.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_-mX18jJkPDk.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_7O9kkDxEvaY.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_MaJlWFemO68.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_hiVs1hNyPpw.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_sdR443ncw-I.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "Two girls are in the kitchen washing dishes. They put the washed dishes into the dishwasher. They start fighting in the kitchen."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_a39_RoOBkX0.mp4",
"activitynet/videos/v_u8ykXBc2Efs.mkv",
"activitynet/videos/v_QryL-hVKAOA.mp4",
"activitynet/videos/v_kuMevlNUDCs.mp4",
"activitynet/videos/v_8iHklV25LaE.mp4"
] | 4
|
<answer>2</answer>
|
[
2
] | false
| 0
| 16,608.478764
| 4,826
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"Two girls are in the kitchen washing dishes. They put the washed dishes into the dishwasher. They start fighting in the kitchen.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_a39_RoOBkX0.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_u8ykXBc2Efs.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_QryL-hVKAOA.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_kuMevlNUDCs.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_8iHklV25LaE.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A man touches a knife on a table. He picks up a sharpener and sharpens the blade of the knife on the sharpener. He demonstrates how much sharper the knife is by cutting paper in between attempts to sharpen it."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_8UXuHMmOYGI.mp4",
"activitynet/videos/v_txyXUXWybt4.mp4",
"activitynet/videos/v_2tf414bkudE.mp4",
"activitynet/videos/v_oA_uJ9gLvUQ.mp4",
"activitynet/videos/v_KU8VVtam3ig.mp4"
] | 3
|
<answer>1</answer>
|
[
1
] | false
| 0
| 14,397.973076
| 4,829
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A man touches a knife on a table. He picks up a sharpener and sharpens the blade of the knife on the sharpener. He demonstrates how much sharper the knife is by cutting paper in between attempts to sharpen it.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_8UXuHMmOYGI.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_txyXUXWybt4.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_2tf414bkudE.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_oA_uJ9gLvUQ.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_KU8VVtam3ig.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A man wearing a green knit cap drives up to the camera on a lawn tractor. The man stops the tractor and gets out of the lawn tractor and walks up to the camera. The camera pans across the yard and takes a close up of a leaf blower laying the yard. A girl in a blue jacket tries to start the leaf blower many times but is unable to. The man in the green knit cap tries twice and the leaf blower starts. The girl in the blue jacket mounts the leaf blower on her back and begins to blow leafs around. The man in the green knit cap is using the leaf blower and the girl in the blue jacket is now driving the lawn tractor."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_GR6Ul2pD8_Y.mp4",
"activitynet/videos/v_8Da6w-Eg3Ko.mp4",
"activitynet/videos/v_2Sev8z4P7pE.mp4",
"activitynet/videos/v_f_BzYUCp4J0.mp4",
"activitynet/videos/v_gY0qgtM0Gt8.mp4"
] | 1
|
<answer>1</answer>
|
[
1
] | true
| 1
| 9,712.934716
| 4,831
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A man wearing a green knit cap drives up to the camera on a lawn tractor. The man stops the tractor and gets out of the lawn tractor and walks up to the camera. The camera pans across the yard and takes a close up of a leaf blower laying the yard. A girl in a blue jacket tries to start the leaf blower many times but is unable to. The man in the green knit cap tries twice and the leaf blower starts. The girl in the blue jacket mounts the leaf blower on her back and begins to blow leafs around. The man in the green knit cap is using the leaf blower and the girl in the blue jacket is now driving the lawn tractor.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_GR6Ul2pD8_Y.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_8Da6w-Eg3Ko.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_2Sev8z4P7pE.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_f_BzYUCp4J0.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_gY0qgtM0Gt8.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A black and white cat sits on a table next to a girl wearing a black and white sweater. The girl holds the cats paw in her left hand. The cat groomer grabs a cat nail clipper in her right hand. The cat groomer squeezes the cats paw and proceeds to clip the cats nails. The cat groomer then holds the cat close and adjust the camera."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_4mzM3JjBJ74.mkv",
"activitynet/videos/v_HcZ3irBAcE0.mp4",
"activitynet/videos/v_mV5DfYFg4H0.mp4",
"activitynet/videos/v_tfPm9xAZ5z0.mp4",
"activitynet/videos/v_Feo8xSjY5A8.mp4"
] | 4
|
<answer>1</answer>
|
[
1
] | false
| 0
| 20,719.424676
| 4,830
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A black and white cat sits on a table next to a girl wearing a black and white sweater. The girl holds the cats paw in her left hand. The cat groomer grabs a cat nail clipper in her right hand. The cat groomer squeezes the cats paw and proceeds to clip the cats nails. The cat groomer then holds the cat close and adjust the camera.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_4mzM3JjBJ74.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_HcZ3irBAcE0.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_mV5DfYFg4H0.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_tfPm9xAZ5z0.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_Feo8xSjY5A8.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A boy is running around a kitchen, dancing as he mops. He appears to be singing, and slides back and forth. The floor is shiny and clean as he works."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_7FPvAakfM9Y.mp4",
"activitynet/videos/v_TbxVdELEiO8.mp4",
"activitynet/videos/v_0w-3O0ZOQFQ.mkv",
"activitynet/videos/v_QylENMzsW9w.mp4",
"activitynet/videos/v_sIzcPVbn0lg.mp4"
] | 1
|
<answer>5</answer>
|
[
5
] | false
| 0
| 11,071.811273
| 4,832
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A boy is running around a kitchen, dancing as he mops. He appears to be singing, and slides back and forth. The floor is shiny and clean as he works.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_7FPvAakfM9Y.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_TbxVdELEiO8.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_0w-3O0ZOQFQ.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_QylENMzsW9w.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_sIzcPVbn0lg.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A man demonstrates how to bowl a ball. He shows the proper technique and form."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_P9qhbSYblG4.mp4",
"activitynet/videos/v_SLdf2ZUdgEQ.mkv",
"activitynet/videos/v_fOuFF7dGPtI.mp4",
"activitynet/videos/v_Qo3riKtRg2c.mp4",
"activitynet/videos/v_JRs2MpyP0SQ.mp4"
] | 3
|
<answer>2</answer>
|
[
2
] | false
| 0
| 17,165.808889
| 4,833
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A man demonstrates how to bowl a ball. He shows the proper technique and form.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_P9qhbSYblG4.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_SLdf2ZUdgEQ.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_fOuFF7dGPtI.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_Qo3riKtRg2c.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_JRs2MpyP0SQ.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "Players are playing a baseball game. A player gives another player a high five. A man in a hat stands on the field. The replay is shown in slow motion. The players are walking off the field."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_079MEwdDNjg.mkv",
"activitynet/videos/v_v5patZyuYys.mp4",
"activitynet/videos/v_we6Ddq1ABcQ.mp4",
"activitynet/videos/v_I2w4N_GnyT8.mp4",
"activitynet/videos/v_qMJi2nXWOkg.mp4"
] | 4
|
<answer>1</answer>
|
[
1
] | false
| 0
| 10,684.395513
| 4,834
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"Players are playing a baseball game. A player gives another player a high five. A man in a hat stands on the field. The replay is shown in slow motion. The players are walking off the field.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_079MEwdDNjg.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_v5patZyuYys.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_we6Ddq1ABcQ.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_I2w4N_GnyT8.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_qMJi2nXWOkg.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A woman takes a swing chair to the side of the backyard. Then, the woman mows the grass of the backyard. After, the woman takes out the container of the mower and puts the dead grass in a trash bag, then she puts back the container in the mower. A boy stands watching the woman."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_5Yq5GMPBguI.mp4",
"activitynet/videos/v_4OIkfJ_IkpA.mkv",
"activitynet/videos/v_62BPME-ikJU.mp4",
"activitynet/videos/v_2Sev8z4P7pE.mp4",
"activitynet/videos/v_5oy5Yi6fzJU.mp4"
] | 3
|
<answer>2</answer>
|
[
2
] | false
| 0
| 9,304.244794
| 4,836
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A woman takes a swing chair to the side of the backyard. Then, the woman mows the grass of the backyard. After, the woman takes out the container of the mower and puts the dead grass in a trash bag, then she puts back the container in the mower. A boy stands watching the woman.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_5Yq5GMPBguI.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_4OIkfJ_IkpA.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_62BPME-ikJU.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_2Sev8z4P7pE.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_5oy5Yi6fzJU.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A girl attempts to do some sort of gymnastic routine on the grass. She does an upside down crab walk across the yard."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_KyDcuYjDi_Y.mkv",
"activitynet/videos/v_62BPME-ikJU.mp4",
"activitynet/videos/v_SyOdA4ZKEtQ.mp4",
"activitynet/videos/v_-VexUX6OJBM.mp4",
"activitynet/videos/v_MMB5Cn3JCGI.mp4"
] | 1
|
<answer>3</answer>
|
[
3
] | false
| 0
| 15,398.491252
| 4,835
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A girl attempts to do some sort of gymnastic routine on the grass. She does an upside down crab walk across the yard.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_KyDcuYjDi_Y.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_62BPME-ikJU.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_SyOdA4ZKEtQ.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_-VexUX6OJBM.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_MMB5Cn3JCGI.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A person is drilling into the ground. They vacuum up what they drilled with a hose. They hit a wood panel into the ground."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_zWiu-wdKeWs.mp4",
"activitynet/videos/v_GP2S0V5NiPs.mp4",
"activitynet/videos/v_YcjLd_XBK5Y.mp4",
"activitynet/videos/v_o0gdMKlKLcU.mp4",
"activitynet/videos/v_CdjU2OZri4c.mp4"
] | 5
|
<answer>2</answer>
|
[
2
] | false
| 0
| 14,100.619152
| 4,837
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A person is drilling into the ground. They vacuum up what they drilled with a hose. They hit a wood panel into the ground.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_zWiu-wdKeWs.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_GP2S0V5NiPs.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_YcjLd_XBK5Y.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_o0gdMKlKLcU.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_CdjU2OZri4c.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A man runs around an area, celebrating and kissing the emblem on his uniform. People are then shown playing soccer in a game. Several goals and interceptions are made, and the teams celebrate."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_o8-v0rPP06U.mp4",
"activitynet/videos/v_q92zSoMudWU.mp4",
"activitynet/videos/v_QoRUUJz-PU0.mp4",
"activitynet/videos/v_8c-s3TKrtdE.mp4",
"activitynet/videos/v_UMhZGJqeSuU.mp4"
] | 3
|
<answer>1</answer>
|
[
1
] | false
| 0
| 11,227.69367
| 4,838
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A man runs around an area, celebrating and kissing the emblem on his uniform. People are then shown playing soccer in a game. Several goals and interceptions are made, and the teams celebrate.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_o8-v0rPP06U.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_q92zSoMudWU.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_QoRUUJz-PU0.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_8c-s3TKrtdE.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_UMhZGJqeSuU.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "How to clean your RV windows and mirrors fast without using any spray. You get a wash cloth and you put it under the faucet to get wet and then you rinse it out so it's not soaking. Then you use it to wash the mirror wrong back and forth and all around. With a dry cloth you go over it and then your mirrors are clean."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_un6VqJYUpDo.mp4",
"activitynet/videos/v_PVdd6E1S0Yc.mp4",
"activitynet/videos/v_tFiXLhbKdnk.mkv",
"activitynet/videos/v_egT7FYHlWho.mp4",
"activitynet/videos/v_ev-RTtbVjFI.mp4"
] | 4
|
<answer>4</answer>
|
[
4
] | true
| 1
| 11,581.806695
| 4,839
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"How to clean your RV windows and mirrors fast without using any spray. You get a wash cloth and you put it under the faucet to get wet and then you rinse it out so it's not soaking. Then you use it to wash the mirror wrong back and forth and all around. With a dry cloth you go over it and then your mirrors are clean.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_un6VqJYUpDo.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_PVdd6E1S0Yc.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_tFiXLhbKdnk.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_egT7FYHlWho.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_ev-RTtbVjFI.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A large group of people are shown from several shots walking around as well as divers jumping off a drive. More shots of athletes and audiences are shown as well as people being interviewed on the camera. More divers jump as well as several other interviews take place."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_WZUxscN9rW8.mp4",
"activitynet/videos/v_M_WEOecjwLY.mkv",
"activitynet/videos/v_5rw6n16ILgY.mp4",
"activitynet/videos/v_gwZleaX_ZR8.mp4",
"activitynet/videos/v_fBlvOzfFq-k.mp4"
] | 5
|
<answer>1</answer>
|
[
1
] | false
| 0
| 16,819.921804
| 4,840
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A large group of people are shown from several shots walking around as well as divers jumping off a drive. More shots of athletes and audiences are shown as well as people being interviewed on the camera. More divers jump as well as several other interviews take place.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_WZUxscN9rW8.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_M_WEOecjwLY.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_5rw6n16ILgY.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_gwZleaX_ZR8.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_fBlvOzfFq-k.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A hand is shown hovering over assorted cutlery. The hand picks up a wooden mallet. The man holding the wooden mallet walks out of the room. The man uses the wooden mallet to hammer down croquet hoops. The man picks up a croquet hammer and a ball. the man sits on a rock with the hammer leaning against him. The man stands up and conjures a clone of himself. The man and his clone engage in a game of croquet. A close up of a croquet mallet falling is shown. A close up of the man's face is shown. The man uses the croquet mallet to shoot a laser at his clown. The man throws his jacket over his clone."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_vGcH8N8sJlM.mp4",
"activitynet/videos/v_AUHORHUgC-s.mp4",
"activitynet/videos/v_IQGg87yZZjs.mp4",
"activitynet/videos/v_gA0m7YUH408.mp4",
"activitynet/videos/v_3q_MOQNfSmA.mp4"
] | 3
|
<answer>3</answer>
|
[
3
] | true
| 1
| 21,405.378544
| 4,841
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A hand is shown hovering over assorted cutlery. The hand picks up a wooden mallet. The man holding the wooden mallet walks out of the room. The man uses the wooden mallet to hammer down croquet hoops. The man picks up a croquet hammer and a ball. the man sits on a rock with the hammer leaning against him. The man stands up and conjures a clone of himself. The man and his clone engage in a game of croquet. A close up of a croquet mallet falling is shown. A close up of the man's face is shown. The man uses the croquet mallet to shoot a laser at his clown. The man throws his jacket over his clone.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_vGcH8N8sJlM.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_AUHORHUgC-s.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_IQGg87yZZjs.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_gA0m7YUH408.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_3q_MOQNfSmA.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "The man in on the beach is flying the kite. The man is using two ropes, he's maneuvering the kite with both hands. The kite is flying round and round, then it flew high then low."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_Y7VWbYGI0Oc.mp4",
"activitynet/videos/v_FEqLmpNzxdg.mp4",
"activitynet/videos/v_aWnpbk007cE.mp4",
"activitynet/videos/v_gXvRxyT5rWw.mp4",
"activitynet/videos/v_ZJ6BFrKcRe0.mp4"
] | 5
|
<answer>1</answer>
|
[
1
] | false
| 0
| 29,131.040297
| 4,842
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"The man in on the beach is flying the kite. The man is using two ropes, he's maneuvering the kite with both hands. The kite is flying round and round, then it flew high then low.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_Y7VWbYGI0Oc.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_FEqLmpNzxdg.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_aWnpbk007cE.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_gXvRxyT5rWw.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_ZJ6BFrKcRe0.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A person is sitting a a mirror putting eye makeup on, alternating from left eye to right eye. They then apply blush with a sponge and brush."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_4OfhHE72V8c.mp4",
"activitynet/videos/v_2HmhRdKRVb4.mp4",
"activitynet/videos/v_B8d9FYuZglQ.mp4",
"activitynet/videos/v_fZQS02Ypca4.mp4",
"activitynet/videos/v_dZZE8HI0OBE.mp4"
] | 1
|
<answer>2</answer>
|
[
2
] | false
| 0
| 26,394.702954
| 4,843
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A person is sitting a a mirror putting eye makeup on, alternating from left eye to right eye. They then apply blush with a sponge and brush.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_4OfhHE72V8c.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_2HmhRdKRVb4.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_B8d9FYuZglQ.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_fZQS02Ypca4.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_dZZE8HI0OBE.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A man at a beach is doing body weight calisthenics on two metal parallel bars. He does leg raises and as his legs become parallel with the bars he does leg splits and then stops. Then he is back on the bars doing the exercise again."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_BJ9r8_JnG0k.mp4",
"activitynet/videos/v_gXvRxyT5rWw.mp4",
"activitynet/videos/v_xVPTVGpOkGE.mp4",
"activitynet/videos/v_94bJbSWNw3o.mp4",
"activitynet/videos/v_r9eXOf4hvCE.mp4"
] | 2
|
<answer>2</answer>
|
[
2
] | true
| 1
| 22,287.017671
| 4,844
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A man at a beach is doing body weight calisthenics on two metal parallel bars. He does leg raises and as his legs become parallel with the bars he does leg splits and then stops. Then he is back on the bars doing the exercise again.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_BJ9r8_JnG0k.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_gXvRxyT5rWw.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_xVPTVGpOkGE.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_94bJbSWNw3o.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_r9eXOf4hvCE.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A man is seen riding a horse onto a field while swinging a rope and a young calf running in front of him. The man then throws the rope onto the calf and ties to animal up then walking away."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_YSnCGTXJtig.mp4",
"activitynet/videos/v_LA5UXJ_hVU4.mp4",
"activitynet/videos/v_gSkE0KCvves.mp4",
"activitynet/videos/v_eUecHAdv1uU.mp4",
"activitynet/videos/v_dKiy-7TZqeI.mp4"
] | 3
|
<answer>2</answer>
|
[
2
] | false
| 0
| 15,543.177661
| 4,845
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A man is seen riding a horse onto a field while swinging a rope and a young calf running in front of him. The man then throws the rope onto the calf and ties to animal up then walking away.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_YSnCGTXJtig.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_LA5UXJ_hVU4.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_gSkE0KCvves.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_eUecHAdv1uU.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_dKiy-7TZqeI.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A man in black suit is burying his backpack in the snow, then he started to ski down the slope, in the forest, jumping on the slope, back flipping, skiing. The skier ski at the side of the water, ski down the sloped, jumping on the edges."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_ZGL-PmMopeM.mp4",
"activitynet/videos/v_7H5oYHs7EJw.mp4",
"activitynet/videos/v_Cc_DmDsXm6M.mp4",
"activitynet/videos/v_zOGg5-Mll4o.mp4",
"activitynet/videos/v_6QrVxwNUbBk.mp4"
] | 1
|
<answer>1</answer>
|
[
1
] | true
| 1
| 19,719.243205
| 4,846
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A man in black suit is burying his backpack in the snow, then he started to ski down the slope, in the forest, jumping on the slope, back flipping, skiing. The skier ski at the side of the water, ski down the sloped, jumping on the edges.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_ZGL-PmMopeM.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_7H5oYHs7EJw.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_Cc_DmDsXm6M.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_zOGg5-Mll4o.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_6QrVxwNUbBk.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A stove is shown with a skillet on an eye with a flame. A chef is then visible and begins stirring a liquid in a pot and cooks something else in another skillet and it eventually begins to catch on fire. Once the fire simmers down,the man throws red sauce on top of the food and he shuffles it around the skillet. The camera then zooms in on the dish as the chef continues to shuffle the food and cook the meal."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_J3DxJ8gI95U.mp4",
"activitynet/videos/v_3j2d27w3x5Q.mp4",
"activitynet/videos/v_Zc8zn0sKfwo.mp4",
"activitynet/videos/v_cIRMaWUTHuk.mp4",
"activitynet/videos/v_bJ5YjjFLGyA.mp4"
] | 1
|
<answer>2</answer>
|
[
2
] | false
| 0
| 26,326.169005
| 4,847
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A stove is shown with a skillet on an eye with a flame. A chef is then visible and begins stirring a liquid in a pot and cooks something else in another skillet and it eventually begins to catch on fire. Once the fire simmers down,the man throws red sauce on top of the food and he shuffles it around the skillet. The camera then zooms in on the dish as the chef continues to shuffle the food and cook the meal.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_J3DxJ8gI95U.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_3j2d27w3x5Q.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_Zc8zn0sKfwo.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_cIRMaWUTHuk.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_bJ5YjjFLGyA.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A female tattoo artist is doing a tattoo back piece for a man at a shop. A title screen appears in black with white letters explaining the video and the tattoo artists name. The tattoo artist is setting up her inks and papers for the session she is about to have. She peels the transfer paper off the client's back revealing the art. In a progression of 12 sessions shows a visual timeline of how the entire back piece for the client comes out as it shows her inking the clients entire back with artwork. There are several still pictures of the client, the tattoo artists and others posing as the client shows the complete tattoo on his back. A black screen with white letters to show the end of the video."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_TjLoGNBzNRA.mp4",
"activitynet/videos/v_KwBuRjh_v9M.mp4",
"activitynet/videos/v_WqnnGmL-lmU.mp4",
"activitynet/videos/v_P3_YQbHXEIs.mp4",
"activitynet/videos/v_KRGiJIHSd9E.mp4"
] | 5
|
<answer>4</answer>
|
[
4
] | false
| 0
| 34,995.574121
| 4,848
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A female tattoo artist is doing a tattoo back piece for a man at a shop. A title screen appears in black with white letters explaining the video and the tattoo artists name. The tattoo artist is setting up her inks and papers for the session she is about to have. She peels the transfer paper off the client's back revealing the art. In a progression of 12 sessions shows a visual timeline of how the entire back piece for the client comes out as it shows her inking the clients entire back with artwork. There are several still pictures of the client, the tattoo artists and others posing as the client shows the complete tattoo on his back. A black screen with white letters to show the end of the video.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_TjLoGNBzNRA.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_KwBuRjh_v9M.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_WqnnGmL-lmU.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_P3_YQbHXEIs.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_KRGiJIHSd9E.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "man is walking in a field in front of a black and red inflatable. two men are standing holding paintball huns in a field with obstacles. man is showing how to holds the gun in a paintball field."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_PJgy8J1f3jg.mp4",
"activitynet/videos/v_aBdrTqSnWbw.mp4",
"activitynet/videos/v_fDPNV463JuE.mp4",
"activitynet/videos/v_-5c9WHk408g.mp4",
"activitynet/videos/v_YSrnHPcdGL4.mp4"
] | 3
|
<answer>2</answer>
|
[
2
] | false
| 0
| 22,178.021525
| 4,849
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"man is walking in a field in front of a black and red inflatable. two men are standing holding paintball huns in a field with obstacles. man is showing how to holds the gun in a paintball field.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_PJgy8J1f3jg.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_aBdrTqSnWbw.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_fDPNV463JuE.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_-5c9WHk408g.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_YSrnHPcdGL4.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "Two people are sitting at a table. They are playing a game of rock paper scissors. There are words shown on the screen."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_Iq9cAZxki9Y.mp4",
"activitynet/videos/v_FTmGHtBdWi0.mp4",
"activitynet/videos/v_A1EflBqBv14.mkv",
"activitynet/videos/v_iaKlx11RAiY.mp4",
"activitynet/videos/v_lkC_md7KKq0.mp4"
] | 1
|
<answer>4</answer>
|
[
4
] | false
| 0
| 18,593.171677
| 4,850
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"Two people are sitting at a table. They are playing a game of rock paper scissors. There are words shown on the screen.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_Iq9cAZxki9Y.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_FTmGHtBdWi0.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_A1EflBqBv14.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_iaKlx11RAiY.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_lkC_md7KKq0.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "Several different people are inside a gym. They are shown riding different kinds of elliptical and ARC trainers at both fast and slow speeds. Some are running, others jumping, then back to the ellipticals."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_pBaeRTgaNBM.mp4",
"activitynet/videos/v_t3eRbi1Uk5E.mp4",
"activitynet/videos/v_arnKDX_ToxE.mp4",
"activitynet/videos/v_6b8h8ztnj9Q.mp4",
"activitynet/videos/v_c6nEk5N4fSU.mp4"
] | 2
|
<answer>2</answer>
|
[
2
] | true
| 1
| 20,634.616591
| 4,851
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"Several different people are inside a gym. They are shown riding different kinds of elliptical and ARC trainers at both fast and slow speeds. Some are running, others jumping, then back to the ellipticals.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_pBaeRTgaNBM.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_t3eRbi1Uk5E.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_arnKDX_ToxE.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_6b8h8ztnj9Q.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_c6nEk5N4fSU.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A man is inside a room in front of a table. He demonstrates how to properly lay a shirt on the table. Then he shows how to use an iron to iron the shirt smoothly."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_fZQclIXmRHE.mp4",
"activitynet/videos/v_kWmf0_XSfBU.mp4",
"activitynet/videos/v_UlWLcqIvLKk.mp4",
"activitynet/videos/v_r-iXUXMP4DY.mp4",
"activitynet/videos/v_EFEI0-awheU.mp4"
] | 4
|
<answer>1</answer>
|
[
1
] | false
| 0
| 23,061.101745
| 4,852
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A man is inside a room in front of a table. He demonstrates how to properly lay a shirt on the table. Then he shows how to use an iron to iron the shirt smoothly.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_fZQclIXmRHE.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_kWmf0_XSfBU.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_UlWLcqIvLKk.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_r-iXUXMP4DY.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_EFEI0-awheU.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "Two people are seen standing on a stage facing one another and leads into them performing a dance routine with one another. People watch the couple dance from the sidelines as the couple continues to move and end by holding a pose."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_1ftLLKrC81s.mkv",
"activitynet/videos/v_dnzcNZBtUG4.mp4",
"activitynet/videos/v_rrKGM5hck1A.mp4",
"activitynet/videos/v_IRvFx8K0gAI.mp4",
"activitynet/videos/v_suyh4tGuScw.mp4"
] | 4
|
<answer>4</answer>
|
[
4
] | true
| 1
| 23,258.224248
| 4,853
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"Two people are seen standing on a stage facing one another and leads into them performing a dance routine with one another. People watch the couple dance from the sidelines as the couple continues to move and end by holding a pose.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_1ftLLKrC81s.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_dnzcNZBtUG4.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_rrKGM5hck1A.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_IRvFx8K0gAI.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_suyh4tGuScw.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "man is vacuuming the seats of a car. man is cleaning the car holding a hosepipe. car enteres in a parking lo and a man is polishing and waxing the car. man is cleaning the leather of the seats. man is cleaning the wheels with a bottle spray."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_XoFikALe8Q0.mp4",
"activitynet/videos/v_Ja8QImLWYII.mp4",
"activitynet/videos/v_kyafh7Ownao.mp4",
"activitynet/videos/v_jmerKGN0VPs.mp4",
"activitynet/videos/v_KNyWPCoHEng.mp4"
] | 5
|
<answer>4</answer>
|
[
4
] | false
| 0
| 11,443.924643
| 4,854
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"man is vacuuming the seats of a car. man is cleaning the car holding a hosepipe. car enteres in a parking lo and a man is polishing and waxing the car. man is cleaning the leather of the seats. man is cleaning the wheels with a bottle spray.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_XoFikALe8Q0.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_Ja8QImLWYII.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_kyafh7Ownao.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_jmerKGN0VPs.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_KNyWPCoHEng.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A man is holding the legs of a man as he does sits ups in a workout room. The man in the background leans forward. The man stops sit ups and starts reaching for the sky as he lays on his back."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_PL1JmxPH7y4.mp4",
"activitynet/videos/v_i2dFL7sGf9c.mp4",
"activitynet/videos/v_fgoXpih2Kws.mp4",
"activitynet/videos/v_XvFv0n2mJUk.mp4",
"activitynet/videos/v_sAi1aMHR89A.mp4"
] | 5
|
<answer>1</answer>
|
[
1
] | false
| 0
| 6,140.832192
| 4,855
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A man is holding the legs of a man as he does sits ups in a workout room. The man in the background leans forward. The man stops sit ups and starts reaching for the sky as he lays on his back.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_PL1JmxPH7y4.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_i2dFL7sGf9c.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_fgoXpih2Kws.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_XvFv0n2mJUk.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_sAi1aMHR89A.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "We see two bright intro screens. We see a man talking near a lady and the lady with different hairstyles. We see the man put a braids in the lades hair and a ponytail. We see the man and the woman with her hair down. The woman puts her hair up in the same style herself. The man and lady smile and wave at the camera. We then see a bright ending screen."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_rXgC48CLncg.mp4",
"activitynet/videos/v_Y8-7fr5bv24.mp4",
"activitynet/videos/v_9AOVI0OCZqg.mp4",
"activitynet/videos/v_ZGK-w7-bkNw.mp4",
"activitynet/videos/v_IBscTNN6qfY.mp4"
] | 2
|
<answer>5</answer>
|
[
5
] | false
| 0
| 17,049.491792
| 4,857
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"We see two bright intro screens. We see a man talking near a lady and the lady with different hairstyles. We see the man put a braids in the lades hair and a ponytail. We see the man and the woman with her hair down. The woman puts her hair up in the same style herself. The man and lady smile and wave at the camera. We then see a bright ending screen.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_rXgC48CLncg.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_Y8-7fr5bv24.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_9AOVI0OCZqg.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_ZGK-w7-bkNw.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_IBscTNN6qfY.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A group of children are playing in a field. They are throwing a white ball to each other in a game of dodgeball. The kids run, trying not to get hit by the ball."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_uID_HFDKFKw.mp4",
"activitynet/videos/v_35SpLMcN-m8.mp4",
"activitynet/videos/v_lQWij22wbNU.mp4",
"activitynet/videos/v_oc4v7GPk05c.mp4",
"activitynet/videos/v_s43eZJ0hy44.mp4"
] | 5
|
<answer>5</answer>
|
[
5
] | true
| 1
| 24,112.399502
| 4,858
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A group of children are playing in a field. They are throwing a white ball to each other in a game of dodgeball. The kids run, trying not to get hit by the ball.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_uID_HFDKFKw.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_35SpLMcN-m8.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_lQWij22wbNU.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_oc4v7GPk05c.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_s43eZJ0hy44.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "We see a man talking and some bike shots. We then see the title screen. The man talks and we see people riding BMX bikes. A series of tips come up on the screen. We then see the ending screen."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_PCoxnf59j5U.mp4",
"activitynet/videos/v_pXcFBfv5Sf4.mp4",
"activitynet/videos/v_RxXBMgsu6uU.mp4",
"activitynet/videos/v_qy-LbstiMYg.mp4",
"activitynet/videos/v_HW5QhCSKTsw.mkv"
] | 2
|
<answer>2</answer>
|
[
2
] | true
| 1
| 26,892.851495
| 4,856
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"We see a man talking and some bike shots. We then see the title screen. The man talks and we see people riding BMX bikes. A series of tips come up on the screen. We then see the ending screen.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_PCoxnf59j5U.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_pXcFBfv5Sf4.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_RxXBMgsu6uU.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_qy-LbstiMYg.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_HW5QhCSKTsw.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A man on a motorcycle is pulling up into a parking space. He gets off the bike and puts his helmet down on it. Then he grabs a rolling bucket with supplies and pulls out a gadget that he uses to start cleaning the window of a house. He uses a raf to wipe down the window and then he is done."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_3pjVV7A6Apw.mp4",
"activitynet/videos/v_Rq1MoqtH8fM.mkv",
"activitynet/videos/v_SBj7yuFEwQI.mp4",
"activitynet/videos/v_mio5dnRbo4w.mkv",
"activitynet/videos/v_LTPrtyWIcA4.mp4"
] | 5
|
<answer>5</answer>
|
[
5
] | true
| 1
| 15,980.394672
| 4,859
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A man on a motorcycle is pulling up into a parking space. He gets off the bike and puts his helmet down on it. Then he grabs a rolling bucket with supplies and pulls out a gadget that he uses to start cleaning the window of a house. He uses a raf to wipe down the window and then he is done.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_3pjVV7A6Apw.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_Rq1MoqtH8fM.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_SBj7yuFEwQI.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_mio5dnRbo4w.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_LTPrtyWIcA4.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A group of people are on a court going around, lifting others and sitting on the back of people like riding a horse when a sign a is given. Then, two people sits on the leg of her playmates. The girls organize to form a human tower, then a boy gives a bouquet of flower to the girl on the top. After, a girl brings a boy with the eyes covered, and when he turns take the girl and holds and hugs her."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_yxcikJ3Hp8w.mp4",
"activitynet/videos/v_JQavlg895jU.mp4",
"activitynet/videos/v_M0sa3xWhFGo.mp4",
"activitynet/videos/v_0Zg-7EgFiC8.mp4",
"activitynet/videos/v_jfhKC2WFDTo.mp4"
] | 4
|
<answer>2</answer>
|
[
2
] | false
| 0
| 20,510.979648
| 4,860
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A group of people are on a court going around, lifting others and sitting on the back of people like riding a horse when a sign a is given. Then, two people sits on the leg of her playmates. The girls organize to form a human tower, then a boy gives a bouquet of flower to the girl on the top. After, a girl brings a boy with the eyes covered, and when he turns take the girl and holds and hugs her.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_yxcikJ3Hp8w.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_JQavlg895jU.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_M0sa3xWhFGo.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_0Zg-7EgFiC8.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_jfhKC2WFDTo.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "The woman is shown sitting on a motorcycle wearing a red bull helmet. He starts the motorcycle and goes away, next the woman stops the motorcycle when she gets near a car and she changes the clothes she has on and gets back on her motorcycle."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_OyKEEws65l8.mp4",
"activitynet/videos/v_Mk9PMED8K4g.mp4",
"activitynet/videos/v_w3N0Pyz2-m0.mp4",
"activitynet/videos/v_K8XNOs0AwaE.mp4",
"activitynet/videos/v_1ebIpLiTCvw.mp4"
] | 2
|
<answer>2</answer>
|
[
2
] | true
| 1
| 23,288.130166
| 4,861
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"The woman is shown sitting on a motorcycle wearing a red bull helmet. He starts the motorcycle and goes away, next the woman stops the motorcycle when she gets near a car and she changes the clothes she has on and gets back on her motorcycle.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_OyKEEws65l8.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_Mk9PMED8K4g.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_w3N0Pyz2-m0.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_K8XNOs0AwaE.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_1ebIpLiTCvw.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A person plays guitar holding on the lap. The woman turns the hand up and tap with the finger, then the woman continues palying. The person A man sit behind the person drinks beer and button his shirt."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_I5F59PkcDWM.mp4",
"activitynet/videos/v_kfwwya1qzXM.mp4",
"activitynet/videos/v_lNvX6h3o4EA.mp4",
"activitynet/videos/v_Y-2nhi8JdO8.mp4",
"activitynet/videos/v_6Ke30NtYOC0.mp4"
] | 1
|
<answer>1</answer>
|
[
1
] | true
| 1
| 26,441.369984
| 4,862
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A person plays guitar holding on the lap. The woman turns the hand up and tap with the finger, then the woman continues palying. The person A man sit behind the person drinks beer and button his shirt.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_I5F59PkcDWM.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_kfwwya1qzXM.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_lNvX6h3o4EA.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_Y-2nhi8JdO8.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_6Ke30NtYOC0.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A person wearing gloves is seen taking items out of a sink and pouring liquid all around the sides. The woman then scrubs the all around as well as the metal bar and around the sink."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_tBNOJJx4Z9k.mp4",
"activitynet/videos/v_x6E92fGgdH4.mkv",
"activitynet/videos/v_dgRYwmcRpuo.mp4",
"activitynet/videos/v_RnShLAifVno.mp4",
"activitynet/videos/v_rmzMfd9ftU8.mp4"
] | 3
|
<answer>3</answer>
|
[
3
] | true
| 1
| 14,601.37369
| 4,863
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A person wearing gloves is seen taking items out of a sink and pouring liquid all around the sides. The woman then scrubs the all around as well as the metal bar and around the sink.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_tBNOJJx4Z9k.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_x6E92fGgdH4.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_dgRYwmcRpuo.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_RnShLAifVno.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_rmzMfd9ftU8.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A young woman bounce a ball on a stick in different places. The young woman trains in a field hockey bouncing and throwing the ball. A man walks close to the goal carrying a backpack. The young woman practice with another player in a field, then she runs. Two teams play field hockey, and the young lady participate in the game. After, the young woman practice alone bouncing a ball."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_LbXhdPZakpo.mp4",
"activitynet/videos/v_6hu3V1PS4vM.mp4",
"activitynet/videos/v_HGduo1zU6Ok.mp4",
"activitynet/videos/v_D5jLypnn6Ps.mp4",
"activitynet/videos/v_TP8lUusp66Y.mp4"
] | 5
|
<answer>4</answer>
|
[
4
] | false
| 0
| 16,996.530026
| 4,864
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A young woman bounce a ball on a stick in different places. The young woman trains in a field hockey bouncing and throwing the ball. A man walks close to the goal carrying a backpack. The young woman practice with another player in a field, then she runs. Two teams play field hockey, and the young lady participate in the game. After, the young woman practice alone bouncing a ball.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_LbXhdPZakpo.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_6hu3V1PS4vM.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_HGduo1zU6Ok.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_D5jLypnn6Ps.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_TP8lUusp66Y.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A black cat with shiny eyes climbs on the lap of a man that holds a nail clipper. Then, the man cuts the nails of the front legs o the cat, then the man pets the cat while holdings the nail clipper."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_s60we-9PBhw.mp4",
"activitynet/videos/v_lETAKUG4pQw.mp4",
"activitynet/videos/v_wy_oDiDK6lk.mp4",
"activitynet/videos/v_tfPm9xAZ5z0.mp4",
"activitynet/videos/v_kh42ufAYMZQ.mp4"
] | 2
|
<answer>2</answer>
|
[
2
] | true
| 1
| 13,623.96379
| 4,865
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A black cat with shiny eyes climbs on the lap of a man that holds a nail clipper. Then, the man cuts the nails of the front legs o the cat, then the man pets the cat while holdings the nail clipper.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_s60we-9PBhw.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_lETAKUG4pQw.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_wy_oDiDK6lk.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_tfPm9xAZ5z0.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_kh42ufAYMZQ.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A large group of people are seen standing around a track with one holding a large pole and running down the track. The man pole volts over the bar and is shown again in slow motion. Several more clips are shown of people pole volting over a beam as well as again in slow motion."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_Hn3-SRXssY4.mp4",
"activitynet/videos/v_ix40OdQd7iE.mp4",
"activitynet/videos/v_19LxLS1_Yn0.mp4",
"activitynet/videos/v_l3EBfLkfAX8.mp4",
"activitynet/videos/v_QwnUZ-5JaOM.mp4"
] | 2
|
<answer>2</answer>
|
[
2
] | true
| 1
| 16,116.590656
| 4,866
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A large group of people are seen standing around a track with one holding a large pole and running down the track. The man pole volts over the bar and is shown again in slow motion. Several more clips are shown of people pole volting over a beam as well as again in slow motion.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_Hn3-SRXssY4.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_ix40OdQd7iE.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_19LxLS1_Yn0.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_l3EBfLkfAX8.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_QwnUZ-5JaOM.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "Very short scenes of athletes training in various sports are shown. A man prepares to play tennis. The man talks to someone off screen. Scenes of the man playing tennis are shown, interspersed with short clips of the man talking, while text overlay explains tennis."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_S5zweEQSnho.mkv",
"activitynet/videos/v_Zt8zZhMs4Es.mp4",
"activitynet/videos/v_akwJwcvfjLA.mp4",
"activitynet/videos/v_Qhs0AjFvcOA.mp4",
"activitynet/videos/v_jDeBuorU4hY.mp4"
] | 4
|
<answer>4</answer>
|
[
4
] | true
| 1
| 16,216.272463
| 4,867
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"Very short scenes of athletes training in various sports are shown. A man prepares to play tennis. The man talks to someone off screen. Scenes of the man playing tennis are shown, interspersed with short clips of the man talking, while text overlay explains tennis.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_S5zweEQSnho.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_Zt8zZhMs4Es.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_akwJwcvfjLA.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_Qhs0AjFvcOA.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_jDeBuorU4hY.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A man wearing a number on his uniform walks to a plate, preparing himself. He swings a disc, warming up. He then throws the disc, and it lands far away in the field. He walks away, then it is shown again in slow motion."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_VxoBV76IkLM.mp4",
"activitynet/videos/v_HYAlS44yzdo.mp4",
"activitynet/videos/v_j6HDZh7W6Z4.mp4",
"activitynet/videos/v_lGESoAdgps8.mp4",
"activitynet/videos/v_vrwJEvpeHyM.mp4"
] | 3
|
<answer>2</answer>
|
[
2
] | false
| 0
| 13,818.321483
| 4,869
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A man wearing a number on his uniform walks to a plate, preparing himself. He swings a disc, warming up. He then throws the disc, and it lands far away in the field. He walks away, then it is shown again in slow motion.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_VxoBV76IkLM.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_HYAlS44yzdo.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_j6HDZh7W6Z4.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_lGESoAdgps8.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_vrwJEvpeHyM.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "We see an opening title screen. We see a lady standing on a track with a score screen showing. The lady then perform her high jump and stands and jumps and pumps her arms. We see a replay of the jump. We then see the ending title screen."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_wtoKUYBw9f4.mp4",
"activitynet/videos/v_sGwra7t-ARo.mp4",
"activitynet/videos/v_MtmQjudesdM.mp4",
"activitynet/videos/v_5rw6n16ILgY.mp4",
"activitynet/videos/v_qlqF8K072UU.mp4"
] | 1
|
<answer>1</answer>
|
[
1
] | true
| 1
| 21,821.884386
| 4,868
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"We see an opening title screen. We see a lady standing on a track with a score screen showing. The lady then perform her high jump and stands and jumps and pumps her arms. We see a replay of the jump. We then see the ending title screen.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_wtoKUYBw9f4.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_sGwra7t-ARo.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_MtmQjudesdM.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_5rw6n16ILgY.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_qlqF8K072UU.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A lady in the driveway below scrapes ice off her windshield with a small scraper. The lady is looking upward but starts to look down as she scrapes."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_15npAlupNU4.mp4",
"activitynet/videos/v_n-BJ753InB0.mp4",
"activitynet/videos/v_YNo7-L8VQWw.mp4",
"activitynet/videos/v_BrgYIg6UXhU.mkv",
"activitynet/videos/v_YCqbvmEG-Uw.mp4"
] | 5
|
<answer>2</answer>
|
[
2
] | false
| 0
| 18,679.454871
| 4,870
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A lady in the driveway below scrapes ice off her windshield with a small scraper. The lady is looking upward but starts to look down as she scrapes.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_15npAlupNU4.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_n-BJ753InB0.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_YNo7-L8VQWw.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_BrgYIg6UXhU.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_YCqbvmEG-Uw.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A close up of a sign is shown followed by the camera capturing many guitars on shelves around him. A man is seen sitting down playing an accordion and pauses to speak to the camera and show off the instrument he is holding. He continues playing more while smiling to the camera and pauses to speak more and hold onto his instrument."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_Vhf-vNRYQEg.mp4",
"activitynet/videos/v_mYfo8LhPB5Y.mp4",
"activitynet/videos/v_HM_rHjh-wqQ.mp4",
"activitynet/videos/v_fKDl_CnA8nY.mp4",
"activitynet/videos/v_ykdRdg1XvFM.mp4"
] | 1
|
<answer>5</answer>
|
[
5
] | false
| 0
| 20,412.840046
| 4,871
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A close up of a sign is shown followed by the camera capturing many guitars on shelves around him. A man is seen sitting down playing an accordion and pauses to speak to the camera and show off the instrument he is holding. He continues playing more while smiling to the camera and pauses to speak more and hold onto his instrument.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_Vhf-vNRYQEg.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_mYfo8LhPB5Y.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_HM_rHjh-wqQ.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_fKDl_CnA8nY.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_ykdRdg1XvFM.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A small group of people are seen swimming around a pool chasing a ball around each other. People are watching on the side as one man throws the ball to another. One catches the ball and continues to play around the pool."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_4HxmQpkryjA.mp4",
"activitynet/videos/v_VrNHEv6aR38.mp4",
"activitynet/videos/v_pYvqbfVY-s8.mp4",
"activitynet/videos/v_OHNH7IV0768.mp4",
"activitynet/videos/v_CL6TbOgnLzA.mp4"
] | 2
|
<answer>2</answer>
|
[
2
] | true
| 1
| 15,258.703175
| 4,872
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A small group of people are seen swimming around a pool chasing a ball around each other. People are watching on the side as one man throws the ball to another. One catches the ball and continues to play around the pool.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_4HxmQpkryjA.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_VrNHEv6aR38.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_pYvqbfVY-s8.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_OHNH7IV0768.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_CL6TbOgnLzA.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "We see the opening title screen. A man in a room holds a bike and talks to the camera. The man adjusts a and talks off the front tire, and folds the bike in half on itself. The man unfolds the bike and puts the tire back on. We see the ending title screen."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_owLv-_CPNJI.mkv",
"activitynet/videos/v_XjV0D7nJx0Q.mp4",
"activitynet/videos/v_6Y8wppTQFPo.mp4",
"activitynet/videos/v_PLqTX6ij52U.mp4",
"activitynet/videos/v_PCoxnf59j5U.mp4"
] | 4
|
<answer>3</answer>
|
[
3
] | false
| 0
| 19,538.507925
| 4,873
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"We see the opening title screen. A man in a room holds a bike and talks to the camera. The man adjusts a and talks off the front tire, and folds the bike in half on itself. The man unfolds the bike and puts the tire back on. We see the ending title screen.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_owLv-_CPNJI.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_XjV0D7nJx0Q.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_6Y8wppTQFPo.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_PLqTX6ij52U.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_PCoxnf59j5U.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A large dog is holding on to a blue leash that is on a smaller dog. He is basically walking her, while the owner walks a third dog. The dog runs around quite a but but then gets back on track following the owner. He stops ahead of the owner to wait for her to begin moving again."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_a2HjLtnVDaY.mp4",
"activitynet/videos/v_rAO-_VxIJng.mp4",
"activitynet/videos/v_MjmDj36sVxM.mp4",
"activitynet/videos/v_Wyr2o0lsSTU.mp4",
"activitynet/videos/v_-Jp86pFKlsw.mp4"
] | 1
|
<answer>1</answer>
|
[
1
] | true
| 1
| 13,323.686633
| 4,874
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A large dog is holding on to a blue leash that is on a smaller dog. He is basically walking her, while the owner walks a third dog. The dog runs around quite a but but then gets back on track following the owner. He stops ahead of the owner to wait for her to begin moving again.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_a2HjLtnVDaY.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_rAO-_VxIJng.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_MjmDj36sVxM.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_Wyr2o0lsSTU.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_-Jp86pFKlsw.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A man is painting on a large black canvas. He flips the painting around."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_Xf0c2abFH3Y.mp4",
"activitynet/videos/v_sEGceBU8icE.mp4",
"activitynet/videos/v_2oNsMva04MM.mp4",
"activitynet/videos/v_4SecbKo1iGE.mp4",
"activitynet/videos/v__8Zk9dfBgPg.mp4"
] | 2
|
<answer>2</answer>
|
[
2
] | true
| 1
| 18,681.957966
| 4,875
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A man is painting on a large black canvas. He flips the painting around.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_Xf0c2abFH3Y.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_sEGceBU8icE.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_2oNsMva04MM.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_4SecbKo1iGE.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v__8Zk9dfBgPg.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A man talks to the camera while another man stands by. The second man maneuvers across monkey bars sideways. The first man resumes talking to the camera. The first man maneuvers across the monkey bars to join the second man. The second man maneuvers across the monkey bars again by taking multiple bars at a time."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_ku65ME0vW8s.mp4",
"activitynet/videos/v_ccfffP3pXrc.mp4",
"activitynet/videos/v_hBjVRKwCUNA.mp4",
"activitynet/videos/v__QTQEw1b_-U.mp4",
"activitynet/videos/v_bHAzuAnnvcU.mp4"
] | 4
|
<answer>3</answer>
|
[
3
] | false
| 0
| 18,599.37696
| 4,876
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A man talks to the camera while another man stands by. The second man maneuvers across monkey bars sideways. The first man resumes talking to the camera. The first man maneuvers across the monkey bars to join the second man. The second man maneuvers across the monkey bars again by taking multiple bars at a time.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_ku65ME0vW8s.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_ccfffP3pXrc.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_hBjVRKwCUNA.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v__QTQEw1b_-U.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_bHAzuAnnvcU.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A series of words are shown on a black screen. A close up of spring leg attachments is shown. Close ups of individuals bouncing while wearing the spring legs are shown. A series of video showing people using the spring legs to perform a variety of tricks in various public environments is shown."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_73AGD3RWPEw.mp4",
"activitynet/videos/v_GNg5kjnJlOE.mp4",
"activitynet/videos/v_M6FdaEstXbI.mp4",
"activitynet/videos/v_dSF2i1OQtMc.mp4",
"activitynet/videos/v_exCENNu1qBU.mp4"
] | 2
|
<answer>2</answer>
|
[
2
] | true
| 1
| 24,224.326566
| 4,877
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A series of words are shown on a black screen. A close up of spring leg attachments is shown. Close ups of individuals bouncing while wearing the spring legs are shown. A series of video showing people using the spring legs to perform a variety of tricks in various public environments is shown.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_73AGD3RWPEw.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_GNg5kjnJlOE.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_M6FdaEstXbI.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_dSF2i1OQtMc.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_exCENNu1qBU.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "Several young men play a game on a Foosball table. A new player arrives, who is a much older man amd plays a game with the younger men. The older man high-fives a younger player."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_esuEWVNHfsM.mp4",
"activitynet/videos/v_UuLBAMSmwgc.mp4",
"activitynet/videos/v_5C0G3BQ-Nds.mp4",
"activitynet/videos/v_l4LFSd-7hxU.mp4",
"activitynet/videos/v_0gLAhptj34w.mp4"
] | 3
|
<answer>2</answer>
|
[
2
] | false
| 0
| 21,409.306807
| 4,878
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"Several young men play a game on a Foosball table. A new player arrives, who is a much older man amd plays a game with the younger men. The older man high-fives a younger player.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_esuEWVNHfsM.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_UuLBAMSmwgc.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_5C0G3BQ-Nds.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_l4LFSd-7hxU.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_0gLAhptj34w.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A girl in a purple shirt enjoys her time at the playground. She goes across the monkey bars effortlessly. As she approaches the other side a young boy is climbing up. She reaches the other side and steps onto the platform. As she steps forward towards another attraction, the boy climbs higher."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_GMHzZXAQzIA.mp4",
"activitynet/videos/v_m16Cn9VA3Lk.mp4",
"activitynet/videos/v_tAEGMVLn0wk.mp4",
"activitynet/videos/v_JXazqQitVdQ.mp4",
"activitynet/videos/v_n2L9F6cMNaE.mkv"
] | 2
|
<answer>2</answer>
|
[
2
] | true
| 1
| 17,527.447741
| 4,879
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A girl in a purple shirt enjoys her time at the playground. She goes across the monkey bars effortlessly. As she approaches the other side a young boy is climbing up. She reaches the other side and steps onto the platform. As she steps forward towards another attraction, the boy climbs higher.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_GMHzZXAQzIA.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_m16Cn9VA3Lk.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_tAEGMVLn0wk.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_JXazqQitVdQ.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_n2L9F6cMNaE.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A man is diving deep into the ocean. He breathes bubbles as he goes through the water. He looks back momentarily at the camera."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_FsS_NCZEfaI.mp4",
"activitynet/videos/v_cHYZPYLwvks.mp4",
"activitynet/videos/v_U5wliityRuU.mp4",
"activitynet/videos/v__WMRdq7yFpA.mp4",
"activitynet/videos/v_7X_wgaRaJYQ.mkv"
] | 4
|
<answer>2</answer>
|
[
2
] | false
| 0
| 20,975.361082
| 4,880
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A man is diving deep into the ocean. He breathes bubbles as he goes through the water. He looks back momentarily at the camera.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_FsS_NCZEfaI.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_cHYZPYLwvks.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_U5wliityRuU.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v__WMRdq7yFpA.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_7X_wgaRaJYQ.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A close up of tools are shown that lead into a person tightening a screw on a bike. The person adds parts to the screw and continuing screwing it around. The man then puts one last screw on and tightens it with a belt."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_4EloxAiCydc.mp4",
"activitynet/videos/v_45llr44Pu9g.mp4",
"activitynet/videos/v_2oizmWFx4PA.mp4",
"activitynet/videos/v_4sm-tTbfamM.mp4",
"activitynet/videos/v_hXSee4C6pyE.mp4"
] | 3
|
<answer>2</answer>
|
[
2
] | false
| 0
| 23,478.24259
| 4,881
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A close up of tools are shown that lead into a person tightening a screw on a bike. The person adds parts to the screw and continuing screwing it around. The man then puts one last screw on and tightens it with a belt.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_4EloxAiCydc.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_45llr44Pu9g.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_2oizmWFx4PA.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_4sm-tTbfamM.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_hXSee4C6pyE.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A close up of ingredients in shown followed by a person cracking an egg and preparing a fryer. The man moves around the pan on the hot stoves followed by putting the egg in the pan. The man cooks the egg and places it on a plate to present."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_3X6eP273RoI.mp4",
"activitynet/videos/v_pwKZRo19Vf0.mp4",
"activitynet/videos/v_ggyGuKFjdxk.mp4",
"activitynet/videos/v_Y9B22Ii7-eE.mp4",
"activitynet/videos/v_4j7sZBThR7s.mp4"
] | 2
|
<answer>2</answer>
|
[
2
] | true
| 1
| 23,748.623407
| 4,882
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A close up of ingredients in shown followed by a person cracking an egg and preparing a fryer. The man moves around the pan on the hot stoves followed by putting the egg in the pan. The man cooks the egg and places it on a plate to present.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_3X6eP273RoI.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_pwKZRo19Vf0.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_ggyGuKFjdxk.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_Y9B22Ii7-eE.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_4j7sZBThR7s.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A gymnast in a pink leotard is standing by getting ready to start her performance for the judges. She gets on the bar and begins her performance, taking her time with her moves doing incredible flips forwards and backwards and landing perfectly on the bar. She is pretty amazing, her continues on very gracefully and beautifully. When she is done with her performance she does a triple flip down off the bar and lands."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_Hi0L9rcsXUI.mp4",
"activitynet/videos/v_c-8GvZKndyQ.mp4",
"activitynet/videos/v_bmWICdhvyJw.mp4",
"activitynet/videos/v_ccKJg_f1UDo.mp4",
"activitynet/videos/v_IJ76Wtgg2g4.mp4"
] | 5
|
<answer>5</answer>
|
[
5
] | true
| 1
| 14,251.341497
| 4,884
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A gymnast in a pink leotard is standing by getting ready to start her performance for the judges. She gets on the bar and begins her performance, taking her time with her moves doing incredible flips forwards and backwards and landing perfectly on the bar. She is pretty amazing, her continues on very gracefully and beautifully. When she is done with her performance she does a triple flip down off the bar and lands.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_Hi0L9rcsXUI.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_c-8GvZKndyQ.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_bmWICdhvyJw.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_ccKJg_f1UDo.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_IJ76Wtgg2g4.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A black elliptical is shown and a woman dressed in red workout clothes gets on it and starts to exercise. The front of the machine is shown and the woman presses a button to speed up the machine and moves faster. Next,she gets off the bike and twists a knob to raise the incline of the machine and continues her workout."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_ESecNZbZgug.mp4",
"activitynet/videos/v_m1aF1CVo-s8.mp4",
"activitynet/videos/v_hc4DBHpRuGM.mp4",
"activitynet/videos/v_mNM01g9wLy4.mp4",
"activitynet/videos/v_cROJALtLB1k.mp4"
] | 1
|
<answer>8</answer>
|
[] | false
| 0
| 19,601.885589
| 4,883
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A black elliptical is shown and a woman dressed in red workout clothes gets on it and starts to exercise. The front of the machine is shown and the woman presses a button to speed up the machine and moves faster. Next,she gets off the bike and twists a knob to raise the incline of the machine and continues her workout.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_ESecNZbZgug.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_m1aF1CVo-s8.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_hc4DBHpRuGM.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_mNM01g9wLy4.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_cROJALtLB1k.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A woman sits on front a red shoe. Then, the woman talks a toothbrush and brushes the red shoe. Next, the woman puts the shoe on the table and continues talking."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_EFGtb9IDQao.mp4",
"activitynet/videos/v_Jsx38_s3Mnc.mkv",
"activitynet/videos/v_kElViDpjunQ.mp4",
"activitynet/videos/v_od1jHUzgrAU.mp4",
"activitynet/videos/v_dFVX_2UQ2WY.mp4"
] | 5
|
<answer>5</answer>
|
[
5
] | true
| 1
| 18,166.529671
| 4,885
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A woman sits on front a red shoe. Then, the woman talks a toothbrush and brushes the red shoe. Next, the woman puts the shoe on the table and continues talking.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_EFGtb9IDQao.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_Jsx38_s3Mnc.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_kElViDpjunQ.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_od1jHUzgrAU.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_dFVX_2UQ2WY.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A large snowy area is seen followed by a person pushing a shovel. The person pushing the snow all around the area and clearing out a walkway. The person continues shoveling while the camera captures them from behind."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_FgRIl4bNl_M.mp4",
"activitynet/videos/v_DBGea9pST1A.mp4",
"activitynet/videos/v_x-rGfBaFQek.mp4",
"activitynet/videos/v_vCeaFAiokrU.mp4",
"activitynet/videos/v_Et4GHTvGbg0.mp4"
] | 1
|
<answer>5</answer>
|
[
5
] | false
| 0
| 16,828.890338
| 4,886
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A large snowy area is seen followed by a person pushing a shovel. The person pushing the snow all around the area and clearing out a walkway. The person continues shoveling while the camera captures them from behind.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_FgRIl4bNl_M.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_DBGea9pST1A.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_x-rGfBaFQek.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_vCeaFAiokrU.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_Et4GHTvGbg0.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A woman sits on a rowing machine an pulls a handlebar with a string. The machine display the pulling force."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_wq4H7L15NMA.mp4",
"activitynet/videos/v_R3HC-IAZVZg.mp4",
"activitynet/videos/v_8-QcL1k5n6k.mp4",
"activitynet/videos/v_ugDN2gDN99E.mp4",
"activitynet/videos/v_cROJALtLB1k.mp4"
] | 5
|
<answer>3</answer>
|
[
3
] | false
| 0
| 17,973.926755
| 4,887
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A woman sits on a rowing machine an pulls a handlebar with a string. The machine display the pulling force.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_wq4H7L15NMA.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_R3HC-IAZVZg.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_8-QcL1k5n6k.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_ugDN2gDN99E.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_cROJALtLB1k.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "Two people are seen working on a roof that leads into clips of a clock and a house. Several shots are shown of a roof as well as the men laying down paper and pushing liquid all over the roof. The men continue this method and finish by laying more paper down."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_xf9iLflgRro.mkv",
"activitynet/videos/v_V9MTU7xLukc.mp4",
"activitynet/videos/v_aPzHheM0Egw.mp4",
"activitynet/videos/v_2VoWT4gnQDg.mkv",
"activitynet/videos/v_4CsTbXdERSU.mp4"
] | 1
|
<answer>2</answer>
|
[
2
] | false
| 0
| 7,269.521712
| 4,888
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"Two people are seen working on a roof that leads into clips of a clock and a house. Several shots are shown of a roof as well as the men laying down paper and pushing liquid all over the roof. The men continue this method and finish by laying more paper down.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_xf9iLflgRro.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_V9MTU7xLukc.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_aPzHheM0Egw.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_2VoWT4gnQDg.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_4CsTbXdERSU.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A man holding a man throws it towards a group of kids. in turn, those kids start throwing many balls back in his direction. The girls on the other side start grabbing the balls and throwing back at the other team and trying not to get hit. The boys start throwing multiple balls at them again."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_-5Q7iNtaWCU.mp4",
"activitynet/videos/v_riuJrZqkYYU.mp4",
"activitynet/videos/v_-erT3ckPkAg.mp4",
"activitynet/videos/v_s43eZJ0hy44.mp4",
"activitynet/videos/v_Wgh8e4V8hBc.mp4"
] | 1
|
<answer>2</answer>
|
[
2
] | false
| 0
| 14,539.072462
| 4,889
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A man holding a man throws it towards a group of kids. in turn, those kids start throwing many balls back in his direction. The girls on the other side start grabbing the balls and throwing back at the other team and trying not to get hit. The boys start throwing multiple balls at them again.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_-5Q7iNtaWCU.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_riuJrZqkYYU.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_-erT3ckPkAg.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_s43eZJ0hy44.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_Wgh8e4V8hBc.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "men are wrapping a gift in a small room. men are playing rugby in a field in the night. men are in baths showein and standing in towels while a man is talking to them. men are going up the stairs and are in a trophy room. woman is talking to men that are in a room and shows them a red gift. men start wraping the gift and using some plastic tape."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_9ZboVy59qrw.mp4",
"activitynet/videos/v_HytB88Fhqw8.mp4",
"activitynet/videos/v_U9pnR51t6As.mp4",
"activitynet/videos/v_mkEME_iWi9o.mp4",
"activitynet/videos/v_juiMCvZUYwk.mp4"
] | 2
|
<answer>2</answer>
|
[
2
] | true
| 1
| 20,325.5889
| 4,890
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"men are wrapping a gift in a small room. men are playing rugby in a field in the night. men are in baths showein and standing in towels while a man is talking to them. men are going up the stairs and are in a trophy room. woman is talking to men that are in a room and shows them a red gift. men start wraping the gift and using some plastic tape.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_9ZboVy59qrw.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_HytB88Fhqw8.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_U9pnR51t6As.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_mkEME_iWi9o.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_juiMCvZUYwk.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "Two bulky men, wearing black top are standing on the table, they put their arms on the table and positioned themselves in an armwrestle match. The men started to arm wrestle while while a referee is watching closely to the men's arm to see who is winning. The men continue to arm wrestle using their all their strength, while the man in black is walking towards the men and take picture, and the men continue to wrestle, while two men in red shirts are doing arm wrestle behind them too, then one of the men won and the crowd cheered."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_1NAlbF88oUI.mp4",
"activitynet/videos/v_2g9GrshWQrU.mp4",
"activitynet/videos/v_9aRUmbcYxUM.mp4",
"activitynet/videos/v_FTmGHtBdWi0.mp4",
"activitynet/videos/v_IaT8-cA_AVU.mp4"
] | 2
|
<answer>4</answer>
|
[
4
] | false
| 0
| 28,834.688106
| 4,891
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"Two bulky men, wearing black top are standing on the table, they put their arms on the table and positioned themselves in an armwrestle match. The men started to arm wrestle while while a referee is watching closely to the men's arm to see who is winning. The men continue to arm wrestle using their all their strength, while the man in black is walking towards the men and take picture, and the men continue to wrestle, while two men in red shirts are doing arm wrestle behind them too, then one of the men won and the crowd cheered.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_1NAlbF88oUI.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_2g9GrshWQrU.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_9aRUmbcYxUM.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_FTmGHtBdWi0.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_IaT8-cA_AVU.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A group of 8 people are wearing helmets and life vests as they are riding a raft in very rough white waters. The raft gets into a bit of trouble,flips over and everyone flips into the very rough waters. The people eventually emerge and they all begin to struggle to get out of the water. One person gets to the other side of the water and is pulled to safety by a rope and someone on land holding the rope. Two people begin to go down the very rough waters as they struggle to keep their heads afloat, and quickly two more people are also going down the rough waters and also struggling to keep their heads afloat. More people are struggling in the water, one is sitting on the upside down raft,and three others are in the rough waters and are also struggling to keep their heads afloat and not bang into rocks."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_2-1MNxfX5Bc.mkv",
"activitynet/videos/v_ByIIq3jFOKo.mkv",
"activitynet/videos/v_cW2R4AuUnK8.mp4",
"activitynet/videos/v_HnCUykqco5M.mp4",
"activitynet/videos/v_gKLbdLKEG6U.mp4"
] | 1
|
<answer>2</answer>
|
[
2
] | false
| 0
| 18,451.472343
| 4,892
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A group of 8 people are wearing helmets and life vests as they are riding a raft in very rough white waters. The raft gets into a bit of trouble,flips over and everyone flips into the very rough waters. The people eventually emerge and they all begin to struggle to get out of the water. One person gets to the other side of the water and is pulled to safety by a rope and someone on land holding the rope. Two people begin to go down the very rough waters as they struggle to keep their heads afloat, and quickly two more people are also going down the rough waters and also struggling to keep their heads afloat. More people are struggling in the water, one is sitting on the upside down raft,and three others are in the rough waters and are also struggling to keep their heads afloat and not bang into rocks.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_2-1MNxfX5Bc.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_ByIIq3jFOKo.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_cW2R4AuUnK8.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_HnCUykqco5M.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_gKLbdLKEG6U.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A man is seen moving a toy airplane around and speaking to another man that walks into frame. The scene leads into clips of several man playing volleyball against one another while many watch on the sides. One of the men then grabs his jacket and leaves."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_HNR_HofJ_Fs.mp4",
"activitynet/videos/v_3FAvxuTw4NI.mp4",
"activitynet/videos/v_CSruNOwxCRY.mp4",
"activitynet/videos/v_CHBpVOfPmRA.mp4",
"activitynet/videos/v_6fWXqCWuU9Y.mp4"
] | 1
|
<answer>2</answer>
|
[
2
] | false
| 0
| 16,705.119647
| 4,893
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A man is seen moving a toy airplane around and speaking to another man that walks into frame. The scene leads into clips of several man playing volleyball against one another while many watch on the sides. One of the men then grabs his jacket and leaves.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_HNR_HofJ_Fs.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_3FAvxuTw4NI.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_CSruNOwxCRY.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_CHBpVOfPmRA.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_6fWXqCWuU9Y.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "An old man talks front a fence and he touches it. Then, the man points the fence on top and down. After, the man continues explainig while touching the fence."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_Z7ZODw0C_hY.mp4",
"activitynet/videos/v_m_gr7WdjJmc.mp4",
"activitynet/videos/v_tyjUDi3uLd0.mp4",
"activitynet/videos/v_2VoWT4gnQDg.mkv",
"activitynet/videos/v_KZyg_UYyL0s.mp4"
] | 5
|
<answer>5</answer>
|
[
5
] | true
| 1
| 13,398.387158
| 4,895
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"An old man talks front a fence and he touches it. Then, the man points the fence on top and down. After, the man continues explainig while touching the fence.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_Z7ZODw0C_hY.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_m_gr7WdjJmc.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_tyjUDi3uLd0.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_2VoWT4gnQDg.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_KZyg_UYyL0s.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "Two men play table tennis while several other individuals watch in the background. A man in red and yellow shorts walks into frame and watches the game. An individual in the back of the room leaves. A man dressed in black walks past the foreground. The man dressed in black briefly reenters the frame. The man in red and yellow shorts walks past the foreground. A man in a white shirt walks past the foreground. A man carrying an object walks past the foreground repeatedly. An individual briefly enters the frame."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_GsR4fagoV-Q.mp4",
"activitynet/videos/v_JnpcJP82WLI.mp4",
"activitynet/videos/v_jDfTrTtPs5s.mp4",
"activitynet/videos/v_pe0MhPhhVIk.mp4",
"activitynet/videos/v_esuEWVNHfsM.mp4"
] | 5
|
<answer>5</answer>
|
[
5
] | true
| 1
| 15,690.953937
| 4,894
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"Two men play table tennis while several other individuals watch in the background. A man in red and yellow shorts walks into frame and watches the game. An individual in the back of the room leaves. A man dressed in black walks past the foreground. The man dressed in black briefly reenters the frame. The man in red and yellow shorts walks past the foreground. A man in a white shirt walks past the foreground. A man carrying an object walks past the foreground repeatedly. An individual briefly enters the frame.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_GsR4fagoV-Q.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_JnpcJP82WLI.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_jDfTrTtPs5s.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_pe0MhPhhVIk.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_esuEWVNHfsM.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A man is seen sitting in a bumper car and begins driving around. Several people are riding around in the cars bumping into one another. The people continue driving around and bumping into one another."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_u2329Chp6IY.mp4",
"activitynet/videos/v_TokZDNwr664.mp4",
"activitynet/videos/v_ZoCRdAYWtKg.mkv",
"activitynet/videos/v_vVvImml1A8g.mp4",
"activitynet/videos/v_mRgS35iyhYE.mp4"
] | 5
|
<answer>2</answer>
|
[
2
] | false
| 0
| 14,247.525684
| 4,896
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A man is seen sitting in a bumper car and begins driving around. Several people are riding around in the cars bumping into one another. The people continue driving around and bumping into one another.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_u2329Chp6IY.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_TokZDNwr664.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_ZoCRdAYWtKg.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_vVvImml1A8g.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_mRgS35iyhYE.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "An ad is shown for a pet wash place. A dog is in a tub, being sprayed with soap before being rubbed down, and finally, rinsed clean. A closing image shows additional information about the wash."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_vgdcVhRSa9E.mp4",
"activitynet/videos/v_6-beYw2R10s.mp4",
"activitynet/videos/v_XKu57UKSqPc.mp4",
"activitynet/videos/v_Xq9ueKle4fY.mp4",
"activitynet/videos/v_WHchTZ61VT4.mp4"
] | 2
|
<answer>2</answer>
|
[
2
] | true
| 1
| 13,911.486332
| 4,897
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"An ad is shown for a pet wash place. A dog is in a tub, being sprayed with soap before being rubbed down, and finally, rinsed clean. A closing image shows additional information about the wash.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_vgdcVhRSa9E.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_6-beYw2R10s.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_XKu57UKSqPc.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_Xq9ueKle4fY.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_WHchTZ61VT4.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "The camera shows a man holding a stick to record himself skating quickly down a long hill. The man continues riding around the area while the camera records him moving at a quick pace."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_1BUnQWRBpYg.mp4",
"activitynet/videos/v_ShKrNPaSdhY.mkv",
"activitynet/videos/v_mYHezmI0U6U.mp4",
"activitynet/videos/v_p8MvTi8hJdE.mp4",
"activitynet/videos/v_mpLYUgMhacA.mp4"
] | 1
|
<answer>2</answer>
|
[
2
] | false
| 0
| 23,788.714597
| 4,898
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"The camera shows a man holding a stick to record himself skating quickly down a long hill. The man continues riding around the area while the camera records him moving at a quick pace.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_1BUnQWRBpYg.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_ShKrNPaSdhY.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_mYHezmI0U6U.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_p8MvTi8hJdE.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_mpLYUgMhacA.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
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.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "We see the opening title screen. We see a man stand in a room talking. The man then begins to demonstrate Capoeira moves slowly. The man then begins to move quickly. The man turns to the right and continues, then turns forward. The man stops and stands upright. A lady joins the man in the room. The lady and the man dance around each other in a circle. The people finish and shake hands. The screen goes black and the ending title loads."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_1imA9vLRd3k.mkv",
"activitynet/videos/v_CoP3xaSZt7A.mkv",
"activitynet/videos/v_WV-Sf5-aCcc.mkv",
"activitynet/videos/v_0e-qdFlRmPU.mkv",
"activitynet/videos/v_D5xp0LuEcKw.mp4"
] | 3
|
<answer>2</answer>
|
[
2
] | false
| 0
| 17,489.079203
| 4,900
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem 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\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"We see the opening title screen. We see a man stand in a room talking. The man then begins to demonstrate Capoeira moves slowly. The man then begins to move quickly. The man turns to the right and continues, then turns forward. The man stops and stands upright. A lady joins the man in the room. The lady and the man dance around each other in a circle. The people finish and shake hands. The screen goes black and the ending title loads.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_1imA9vLRd3k.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_CoP3xaSZt7A.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_WV-Sf5-aCcc.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_0e-qdFlRmPU.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_D5xp0LuEcKw.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.