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: "A person is skating on a tennis court. They wind their feet in and out of mini cones as they go. They repeat the act in city parking lots."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_tpDhYD9e_cU.mp4",
"activitynet/videos/v_mzVJHw9Jrb4.mp4",
"activitynet/videos/v_exCENNu1qBU.mp4",
"activitynet/videos/v_4T8uFygBeNo.mkv",
"activitynet/videos/v_rbKPBMRj9jY.mp4"
] | 1
|
<answer>2</answer>
|
[
2
] | false
| 0
| 26,733.445809
| 99
|
[
{
"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 skating on a tennis court. They wind their feet in and out of mini cones as they go. They repeat the act in city parking lots.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_tpDhYD9e_cU.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_mzVJHw9Jrb4.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_exCENNu1qBU.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_4T8uFygBeNo.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_rbKPBMRj9jY.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 matador got knocked down by a bull and the bull injured him. Another matador helped him get away after he was gored in the neck. A matador in a gold costume waved a red banner at him ,but he avoided the bull. Several other matadors came out and held up pink banners around the bull then the bull fell to the ground."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_FQVs9_IbgOY.mp4",
"activitynet/videos/v_Qnr73D2zIjU.mp4",
"activitynet/videos/v_SKdouCRLoKE.mp4",
"activitynet/videos/v_Aa33vHLEXJA.mp4",
"activitynet/videos/v_feWO_gqAcGk.mp4"
] | 4
|
<answer>3</answer>
|
[
3
] | false
| 0
| 10,068.923248
| 102
|
[
{
"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 matador got knocked down by a bull and the bull injured him. Another matador helped him get away after he was gored in the neck. A matador in a gold costume waved a red banner at him ,but he avoided the bull. Several other matadors came out and held up pink banners around the bull then the bull fell to 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_FQVs9_IbgOY.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_Qnr73D2zIjU.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_SKdouCRLoKE.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_Aa33vHLEXJA.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_feWO_gqAcGk.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: "This young black woman is peeling the onion over the garbage can in the kitchen. There is a small dog standing next to the garbage can, but the camera turns so the dog isn't shown. Then someone walks in while holding flamin' hot funyun chips in their hand."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_L_8Gyi8FMk4.mp4",
"activitynet/videos/v_ave_VDl3LwE.mp4",
"activitynet/videos/v_lgu-DBDWlEs.mp4",
"activitynet/videos/v_0r-_a6m5k-0.mp4",
"activitynet/videos/v_64UBH371Jj8.mp4"
] | 3
|
<answer>3</answer>
|
[
3
] | true
| 1
| 20,101.798926
| 101
|
[
{
"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: \"This young black woman is peeling the onion over the garbage can in the kitchen. There is a small dog standing next to the garbage can, but the camera turns so the dog isn't shown. Then someone walks in while holding flamin' hot funyun chips in their hand.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_L_8Gyi8FMk4.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_ave_VDl3LwE.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_lgu-DBDWlEs.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_0r-_a6m5k-0.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_64UBH371Jj8.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 elderly man walks to a towel that is spread on his living room floor. The man lies down on the towel, lifts his bent legs and begins doing a lot of sit ups where he touches his elbows to his knees. The man finally stops, kicks himself up to a sitting position, stands up and walks towards the side of the camera."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_CdwgIN9FkdY.mp4",
"activitynet/videos/v_Sul7NDmB5HM.mp4",
"activitynet/videos/v_JOBSEatasv4.mp4",
"activitynet/videos/v_37pnsj0hlZ4.mp4",
"activitynet/videos/v_je5KvCND9xo.mp4"
] | 3
|
<answer>3</answer>
|
[
3
] | true
| 1
| 15,410.904754
| 104
|
[
{
"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 elderly man walks to a towel that is spread on his living room floor. The man lies down on the towel, lifts his bent legs and begins doing a lot of sit ups where he touches his elbows to his knees. The man finally stops, kicks himself up to a sitting position, stands up and walks towards the side of 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_CdwgIN9FkdY.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_Sul7NDmB5HM.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_JOBSEatasv4.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_37pnsj0hlZ4.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_je5KvCND9xo.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 is riding on the side of the car, the car is going side to side. The car is running on one side while the other side is up. Five men went out of the the car and began removing the tires, then they cheered when they removed the tires. The men put back the tires while the car is still running on one side."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_GKBYgS99oiI.mp4",
"activitynet/videos/v_cIpBpGQ0XTI.mp4",
"activitynet/videos/v_B4qwjeJBk0s.mp4",
"activitynet/videos/v_xIhTY02lRSE.mp4",
"activitynet/videos/v_PJgB6h-fImY.mp4"
] | 2
|
<answer>2</answer>
|
[
2
] | true
| 1
| 22,773.566935
| 103
|
[
{
"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 is riding on the side of the car, the car is going side to side. The car is running on one side while the other side is up. Five men went out of the the car and began removing the tires, then they cheered when they removed the tires. The men put back the tires while the car is still running on one side.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_GKBYgS99oiI.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_cIpBpGQ0XTI.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_B4qwjeJBk0s.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_xIhTY02lRSE.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_PJgB6h-fImY.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 male gymnast is on a mat in front of judges preparing to begin his routine. The boy then jumps on the beam grabbing the bars and doing several spins across the balance beam. He then moves into a hand stand and jumps off the bar into the floor."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_mk3srKjFB3A.mp4",
"activitynet/videos/v_eSpPY2yMg70.mp4",
"activitynet/videos/v_pOVICBn8QMw.mkv",
"activitynet/videos/v_otWTm1_aAqI.mp4",
"activitynet/videos/v_8BsIeOSzK_U.mp4"
] | 4
|
<answer>1</answer>
|
[
1
] | false
| 0
| 25,842.979827
| 105
|
[
{
"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 male gymnast is on a mat in front of judges preparing to begin his routine. The boy then jumps on the beam grabbing the bars and doing several spins across the balance beam. He then moves into a hand stand and jumps off the bar into the floor.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_mk3srKjFB3A.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_eSpPY2yMg70.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_pOVICBn8QMw.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_otWTm1_aAqI.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_8BsIeOSzK_U.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 standing in a dark room weding a metal surface. man is wearing a helmet welding in a suface. welder is holding a welding machine."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_PEpfA3L4m20.mkv",
"activitynet/videos/v_4x0LdQRN248.mp4",
"activitynet/videos/v_XQaaA2UZYh8.mp4",
"activitynet/videos/v_Czw85LWCGes.mp4",
"activitynet/videos/v_pPM1jC_NlzI.mp4"
] | 1
|
<answer>2</answer>
|
[
2
] | false
| 0
| 18,341.549967
| 106
|
[
{
"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 standing in a dark room weding a metal surface. man is wearing a helmet welding in a suface. welder is holding a welding machine.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_PEpfA3L4m20.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_4x0LdQRN248.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_XQaaA2UZYh8.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_Czw85LWCGes.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_pPM1jC_NlzI.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 women are running down a sidewalk. They walk into a house and look at a calendar before going to shave their legs and get ready for a date. The date has flowers when the woman opens the door, but the other woman is not yet done, because her shaving cream took too long."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_gwKy0W1xof4.mp4",
"activitynet/videos/v_gh0GD6OvLHE.mp4",
"activitynet/videos/v_t7J7SugZPlE.mp4",
"activitynet/videos/v_9XjHgUP5QW0.mp4",
"activitynet/videos/v_TaLEPzEyZ34.mp4"
] | 4
|
<answer>4</answer>
|
[
4
] | true
| 1
| 7,194.928327
| 108
|
[
{
"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 women are running down a sidewalk. They walk into a house and look at a calendar before going to shave their legs and get ready for a date. The date has flowers when the woman opens the door, but the other woman is not yet done, because her shaving cream took too long.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_gwKy0W1xof4.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_gh0GD6OvLHE.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_t7J7SugZPlE.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_9XjHgUP5QW0.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_TaLEPzEyZ34.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 waterski in a body of water using a boat as propulsion in the water. A man gets on a water ski with both feet and begins to waterski while holding on to a rope attached to a boat. The camera shows a reverse play of a person doing a backflip out of the water an onto water platform attached to a boat. The people continue to water ski in narrow calm water surrounded by foliage."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_LMMimz1-fa0.mp4",
"activitynet/videos/v_JoQywfQ6B-8.mkv",
"activitynet/videos/v_iId8WcbiKZI.mp4",
"activitynet/videos/v_bYUmtLBL7W4.mkv",
"activitynet/videos/v_vbHLA5l_BRc.mkv"
] | 4
|
<answer>2</answer>
|
[
2
] | false
| 0
| 26,004.854472
| 107
|
[
{
"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 waterski in a body of water using a boat as propulsion in the water. A man gets on a water ski with both feet and begins to waterski while holding on to a rope attached to a boat. The camera shows a reverse play of a person doing a backflip out of the water an onto water platform attached to a boat. The people continue to water ski in narrow calm water surrounded by foliage.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_LMMimz1-fa0.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_JoQywfQ6B-8.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_iId8WcbiKZI.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_bYUmtLBL7W4.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_vbHLA5l_BRc.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 blue opening screen. A man walks out into his yard carrying a shovel. The man then puts on his gloves and dances a bit in the yard. The man shovels the snow. The man spins the shovel in the air dances and shovels again. The man spins the shovel again and words appear on the screen and he shovels more snow. The man walks back into his garage. We see the blue ending title screen."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_VSeBb4e9ysU.mp4",
"activitynet/videos/v_olBh9KMAHMQ.mp4",
"activitynet/videos/v_x-rGfBaFQek.mp4",
"activitynet/videos/v_bb9AIdvKkZU.mp4",
"activitynet/videos/v_RD7AUdgtchE.mp4"
] | 2
|
<answer>2</answer>
|
[
2
] | true
| 1
| 17,010.279943
| 109
|
[
{
"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 blue opening screen. A man walks out into his yard carrying a shovel. The man then puts on his gloves and dances a bit in the yard. The man shovels the snow. The man spins the shovel in the air dances and shovels again. The man spins the shovel again and words appear on the screen and he shovels more snow. The man walks back into his garage. We see the blue 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_VSeBb4e9ysU.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_olBh9KMAHMQ.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_bb9AIdvKkZU.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_RD7AUdgtchE.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 sitting in bumper cars in the middle of a large crowd. The cars then drive around one another and bump into each other while the people watch around the sides."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_nX-GvQmf5Tc.mp4",
"activitynet/videos/v_GgiaxJ1JeSM.mp4",
"activitynet/videos/v_EVtM8DKW4bc.mp4",
"activitynet/videos/v_TokZDNwr664.mp4",
"activitynet/videos/v__crwKCjKRjg.mp4"
] | 3
|
<answer>1</answer>
|
[
1
] | false
| 0
| 17,414.525369
| 110
|
[
{
"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 sitting in bumper cars in the middle of a large crowd. The cars then drive around one another and bump into each other while the people watch around the sides.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_nX-GvQmf5Tc.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_GgiaxJ1JeSM.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_EVtM8DKW4bc.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_TokZDNwr664.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__crwKCjKRjg.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 indoor track meet is shown with the triple jump. Several men complete a triple jump and get their scores."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_U0qUFAPUg_Y.mp4",
"activitynet/videos/v_qlqF8K072UU.mp4",
"activitynet/videos/v_VvD2fdPNWEg.mp4",
"activitynet/videos/v_QwnUZ-5JaOM.mp4",
"activitynet/videos/v_B53aHHzgTzc.mkv"
] | 3
|
<answer>2</answer>
|
[
2
] | false
| 0
| 19,964.890797
| 111
|
[
{
"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 indoor track meet is shown with the triple jump. Several men complete a triple jump and get their scores.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_U0qUFAPUg_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_qlqF8K072UU.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_VvD2fdPNWEg.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_QwnUZ-5JaOM.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_B53aHHzgTzc.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 boy walks into the bathroom at home. The boy lifts up a cup of mouthwash and pours it into his mouth. The boy spits it back out in the sink."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_iSJ87SnNLPc.mp4",
"activitynet/videos/v_s9MNW35YCMw.mp4",
"activitynet/videos/v_Tc0nHNkf0KM.mp4",
"activitynet/videos/v_81dGQTVec_s.mp4",
"activitynet/videos/v_KFS_lGlO-Ew.mp4"
] | 5
|
<answer>2</answer>
|
[
2
] | false
| 0
| 17,660.073725
| 112
|
[
{
"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 boy walks into the bathroom at home. The boy lifts up a cup of mouthwash and pours it into his mouth. The boy spits it back out in 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_iSJ87SnNLPc.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_s9MNW35YCMw.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_Tc0nHNkf0KM.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_81dGQTVec_s.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_KFS_lGlO-Ew.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 boat floats in the ocean and we see divers underwater and sealife. A diver waves and gives a thumbs up and another waves to the camera. A baby shark swims along the ocean floor. We see a series of diver shots and a manta ray swims on the ocean floor. People wave to the camera underwater. 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_ZcgahXg_ELw.mp4",
"activitynet/videos/v_t-8wEopB3AQ.mp4",
"activitynet/videos/v_FsS_NCZEfaI.mp4",
"activitynet/videos/v_U5wliityRuU.mp4",
"activitynet/videos/v_P1gGM89_T2g.mkv"
] | 1
|
<answer>1</answer>
|
[
1
] | true
| 1
| 23,458.491186
| 113
|
[
{
"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 boat floats in the ocean and we see divers underwater and sealife. A diver waves and gives a thumbs up and another waves to the camera. A baby shark swims along the ocean floor. We see a series of diver shots and a manta ray swims on the ocean floor. People wave to the camera underwater. 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_ZcgahXg_ELw.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_t-8wEopB3AQ.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_FsS_NCZEfaI.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_U5wliityRuU.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_P1gGM89_T2g.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 speaking to the camera and pans around an old table. The girl is then seen outside sanding down the table and leads into her painting the table. Several shots of the table are shown as well as lists drawn out, paper being rolled, and laid out onto the table. Finally the table is complete and is shown again inside next to a dress while the woman finishes speaking."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_4ImpZRtbzYw.mp4",
"activitynet/videos/v_zhH7wxXrGSY.mp4",
"activitynet/videos/v__3xMhj4mbsk.mp4",
"activitynet/videos/v_8zq6C0SRyDQ.mp4",
"activitynet/videos/v_I9NukwdINyY.mp4"
] | 1
|
<answer>2</answer>
|
[
2
] | false
| 0
| 17,248.357946
| 114
|
[
{
"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 speaking to the camera and pans around an old table. The girl is then seen outside sanding down the table and leads into her painting the table. Several shots of the table are shown as well as lists drawn out, paper being rolled, and laid out onto the table. Finally the table is complete and is shown again inside next to a dress while the woman finishes speaking.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_4ImpZRtbzYw.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_zhH7wxXrGSY.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__3xMhj4mbsk.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_8zq6C0SRyDQ.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_I9NukwdINyY.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 seen sitting in front of the camera speaking and playing with one another. The girls hold up glasses of mouth wash as well as the bottle itself and begin drinking the liquid. The two swish it around their mouths and spit it out in the end while laughing to the camera."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_tZswexUR6Q0.mkv",
"activitynet/videos/v_wD-9KvI1-AI.mp4",
"activitynet/videos/v_V0e5tItt1RM.mp4",
"activitynet/videos/v_ERGoTBC8NkA.mp4",
"activitynet/videos/v_gA7GpvB10UY.mkv"
] | 2
|
<answer>2</answer>
|
[
2
] | true
| 1
| 16,369.075917
| 115
|
[
{
"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 seen sitting in front of the camera speaking and playing with one another. The girls hold up glasses of mouth wash as well as the bottle itself and begin drinking the liquid. The two swish it around their mouths and spit it out in the end while laughing to 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_tZswexUR6Q0.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_wD-9KvI1-AI.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_V0e5tItt1RM.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_ERGoTBC8NkA.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_gA7GpvB10UY.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 leg is seen followed by a large razor. The person then begins using the razor on their leg up and down. The person continues to use the razor on their leg."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_Jv-bPV8eswU.mp4",
"activitynet/videos/v_0fsMeZoZzJI.mp4",
"activitynet/videos/v_EolA3Rd_Vm4.mp4",
"activitynet/videos/v_5JCqKshcfHE.mp4",
"activitynet/videos/v_AO-0r8H2DOo.mp4"
] | 5
|
<answer>4</answer>
|
[
4
] | false
| 0
| 15,073.191749
| 116
|
[
{
"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 leg is seen followed by a large razor. The person then begins using the razor on their leg up and down. The person continues to use the razor on their leg.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_Jv-bPV8eswU.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_0fsMeZoZzJI.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_EolA3Rd_Vm4.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_5JCqKshcfHE.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_AO-0r8H2DOo.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 standing with a tied up dog outside and begins scrubbing the dog down with soap. The woman moves the dog a bit while scrubbing and shows another dog looking at the camera. The camera pans back to the woman who is still scrubbing the dog."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_FmLxc-aNh88.mkv",
"activitynet/videos/v_JDM9Akcs96g.mp4",
"activitynet/videos/v_Xq9ueKle4fY.mp4",
"activitynet/videos/v_WHchTZ61VT4.mp4",
"activitynet/videos/v_DEduSDgovOQ.mp4"
] | 4
|
<answer>2</answer>
|
[
2
] | false
| 0
| 10,516.291941
| 118
|
[
{
"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 standing with a tied up dog outside and begins scrubbing the dog down with soap. The woman moves the dog a bit while scrubbing and shows another dog looking at the camera. The camera pans back to the woman who is still scrubbing the dog.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_FmLxc-aNh88.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_JDM9Akcs96g.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_Xq9ueKle4fY.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_WHchTZ61VT4.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_DEduSDgovOQ.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 shots are shown of a man speaking to various groups of people and leads into people wearing wet suits and walking. The people walk down a beach and are seen swimming around in the water. More shots of fish are shown and ends with the people walking out of the water and high fiving the camera man."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v__R-jA9hOFCE.mp4",
"activitynet/videos/v_VSONGdnvKiM.mp4",
"activitynet/videos/v_U5wliityRuU.mp4",
"activitynet/videos/v_SibfKtVX3CQ.mp4",
"activitynet/videos/v_IlKOWIBAEFE.mp4"
] | 5
|
<answer>2</answer>
|
[
2
] | false
| 0
| 22,044.985136
| 117
|
[
{
"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 shots are shown of a man speaking to various groups of people and leads into people wearing wet suits and walking. The people walk down a beach and are seen swimming around in the water. More shots of fish are shown and ends with the people walking out of the water and high fiving the camera man.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v__R-jA9hOFCE.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_VSONGdnvKiM.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_SibfKtVX3CQ.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_IlKOWIBAEFE.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 plays a harp in a large church. Another woman plays a flute in the large church."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_r1mrueEHDDE.mp4",
"activitynet/videos/v_z-ttrQ38mOc.mp4",
"activitynet/videos/v_cFCN9QE1M0c.mp4",
"activitynet/videos/v_p42wxuN8MZE.mp4",
"activitynet/videos/v_0RIc6mwDRaQ.mp4"
] | 5
|
<answer>5</answer>
|
[
5
] | true
| 1
| 24,316.000838
| 119
|
[
{
"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 plays a harp in a large church. Another woman plays a flute in the large church.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_r1mrueEHDDE.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_z-ttrQ38mOc.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_cFCN9QE1M0c.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_p42wxuN8MZE.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_0RIc6mwDRaQ.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 child is seen riding down a slide and running to the top again. The boy slides down several more times while other children around him play."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_TxYZLJQOHvY.mp4",
"activitynet/videos/v_2SYTRqm4Ym4.mp4",
"activitynet/videos/v_Br1Ty6PCrv8.mp4",
"activitynet/videos/v_4Lu8ECLHvK4.mp4",
"activitynet/videos/v_q2KR3lzTrq4.mkv"
] | 3
|
<answer>1</answer>
|
[
1
] | false
| 0
| 12,778.19074
| 121
|
[
{
"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 child is seen riding down a slide and running to the top again. The boy slides down several more times while other children around him play.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_TxYZLJQOHvY.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_2SYTRqm4Ym4.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_Br1Ty6PCrv8.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_4Lu8ECLHvK4.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_q2KR3lzTrq4.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 are sitting in tubes on a river. A green tube is shown in the water. A woman wearing a white hat is sitting in a tube."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_yJSQmNSFlNI.mkv",
"activitynet/videos/v_akUXL2VzFEs.mp4",
"activitynet/videos/v_Qu-Y2u1Xn_U.mp4",
"activitynet/videos/v_3nX5ZwzHftM.mp4",
"activitynet/videos/v_dFgwKTH-FhY.mp4"
] | 4
|
<answer>1</answer>
|
[
1
] | false
| 0
| 23,692.294387
| 120
|
[
{
"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 are sitting in tubes on a river. A green tube is shown in the water. A woman wearing a white hat is sitting in a tube.\"\n\nCandidates:\n[1] 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[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_akUXL2VzFEs.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_Qu-Y2u1Xn_U.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_3nX5ZwzHftM.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_dFgwKTH-FhY.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 bull appears in a large rodeo ring. Clowns taunt the bull into charging them and other items. The bull leaves the arena after chasing off a clown."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_Cg_jN5G1ZpY.mp4",
"activitynet/videos/v_PBzlHfEMU5s.mp4",
"activitynet/videos/v_vBKIXqRd-eA.mkv",
"activitynet/videos/v_w0d32MVTY9Q.mp4",
"activitynet/videos/v_BFxxrjqgF0w.mp4"
] | 1
|
<answer>5</answer>
|
[
5
] | false
| 0
| 13,169.65776
| 124
|
[
{
"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 bull appears in a large rodeo ring. Clowns taunt the bull into charging them and other items. The bull leaves the arena after chasing off a clown.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_Cg_jN5G1ZpY.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_PBzlHfEMU5s.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_vBKIXqRd-eA.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_w0d32MVTY9Q.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_BFxxrjqgF0w.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 men are outside in a large body of water,holding on to large sails as the move on their surfboard. The tasks continues,and the men pass more people and start to make circles around each other."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_gdyEfPbUEjw.mp4",
"activitynet/videos/v_GV_BDNmUiLY.mp4",
"activitynet/videos/v_rdkPwRWW91s.mp4",
"activitynet/videos/v_CqscMsSNiNY.mp4",
"activitynet/videos/v_s5y4xXcphcc.mp4"
] | 5
|
<answer>3</answer>
|
[
3
] | false
| 0
| 21,544.884784
| 123
|
[
{
"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 men are outside in a large body of water,holding on to large sails as the move on their surfboard. The tasks continues,and the men pass more people and start to make circles around each other.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_gdyEfPbUEjw.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_GV_BDNmUiLY.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_rdkPwRWW91s.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_CqscMsSNiNY.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_s5y4xXcphcc.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 a blue jacket is skiing down a hill. He sets up a camera in the snow. People are riding a ski lift up the mountain."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_eaI8My4pGq4.mp4",
"activitynet/videos/v_lQUqzLT7bl0.mkv",
"activitynet/videos/v_UjJ8yWaFNGg.mp4",
"activitynet/videos/v_vBpYwyXfE0o.mp4",
"activitynet/videos/v__QyQSAtMdj8.mp4"
] | 3
|
<answer>2</answer>
|
[
2
] | false
| 0
| 31,860.126249
| 122
|
[
{
"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 a blue jacket is skiing down a hill. He sets up a camera in the snow. People are riding a ski lift up the mountain.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_eaI8My4pGq4.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_lQUqzLT7bl0.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_UjJ8yWaFNGg.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_vBpYwyXfE0o.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__QyQSAtMdj8.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 standing in a basket court practicing free throws. man with white shirt is practicing in the background. man stnanding in court is talking to the camera and explaining how to play."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_BwR1DPCVsP8.mp4",
"activitynet/videos/v_0GWJ-VHFlTk.mp4",
"activitynet/videos/v_UfjR8ewF8xo.mp4",
"activitynet/videos/v_ukPz_13Agis.mp4",
"activitynet/videos/v_Mpph0kFsyZ0.mp4"
] | 3
|
<answer>2</answer>
|
[
2
] | false
| 0
| 18,531.956227
| 125
|
[
{
"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 standing in a basket court practicing free throws. man with white shirt is practicing in the background. man stnanding in court is talking to the camera and explaining how to play.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_BwR1DPCVsP8.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_0GWJ-VHFlTk.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_UfjR8ewF8xo.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_ukPz_13Agis.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_Mpph0kFsyZ0.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: "Various pictures of javelin athletes are shown that lead into videos of the athletes throwing the pole. A man then steps up and demonstrates how to properly throw the pole while bouncing up and down."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_9Zy5ylJYiA4.mp4",
"activitynet/videos/v_9-yueOtwiL8.mp4",
"activitynet/videos/v_j7hSNqcWIO4.mp4",
"activitynet/videos/v_V9LudLaWGOM.mkv",
"activitynet/videos/v_ix40OdQd7iE.mp4"
] | 4
|
<answer>4</answer>
|
[
4
] | true
| 1
| 16,829.461088
| 126
|
[
{
"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: \"Various pictures of javelin athletes are shown that lead into videos of the athletes throwing the pole. A man then steps up and demonstrates how to properly throw the pole while bouncing up and down.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_9Zy5ylJYiA4.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_9-yueOtwiL8.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_j7hSNqcWIO4.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_V9LudLaWGOM.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_ix40OdQd7iE.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 couple of arm wrestlers are on a mat. They lock arms and push hard. They keep fighting until one of them wins."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_FTmGHtBdWi0.mp4",
"activitynet/videos/v_ywsH9kD033I.mp4",
"activitynet/videos/v_i4SNM6xSLI8.mp4",
"activitynet/videos/v_bDK-_jU_KzI.mp4",
"activitynet/videos/v_a_uamUiKq1o.mp4"
] | 3
|
<answer>3</answer>
|
[
3
] | true
| 1
| 21,065.578523
| 127
|
[
{
"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 couple of arm wrestlers are on a mat. They lock arms and push hard. They keep fighting until one of them wins.\"\n\nCandidates:\n[1] 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[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_ywsH9kD033I.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_i4SNM6xSLI8.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_bDK-_jU_KzI.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_a_uamUiKq1o.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 men and women are in a pool playing volleyball without the net. The man continues holding the ball up talking to the people and then finally throws the ball. Another standstill takes place and another person swims to the ball but the ball is thrown. They continue throwing the ball back and forth to one another."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_OHNH7IV0768.mp4",
"activitynet/videos/v_esTcWwmykKQ.mp4",
"activitynet/videos/v_LM8C4FSpN0w.mp4",
"activitynet/videos/v_7p99ez6MEeo.mp4",
"activitynet/videos/v_Uc1_7BXtXZs.mp4"
] | 2
|
<answer>5</answer>
|
[
5
] | false
| 0
| 17,416.036776
| 128
|
[
{
"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 men and women are in a pool playing volleyball without the net. The man continues holding the ball up talking to the people and then finally throws the ball. Another standstill takes place and another person swims to the ball but the ball is thrown. They continue throwing the ball back and forth to 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_OHNH7IV0768.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_esTcWwmykKQ.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_LM8C4FSpN0w.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_7p99ez6MEeo.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_Uc1_7BXtXZs.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 sits in a canoe that is docked in shallow water and talks to a camera with the paddle resting across the boat in front of him as a small black dog in a life jacket walks around the boat. The man reaches to pet the dog as the dog walks around and then picks the dog up and holds the dog up by a strap on the top of the life jacket over the water and then sets the dog down back into the boat by the life jacket. The man begins to paddle the boat out into the water as the dog watches on from the side of the boat."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_B0cb0B90Ubg.mp4",
"activitynet/videos/v_WqDep-4l0yc.mp4",
"activitynet/videos/v_AZn294ubbps.mp4",
"activitynet/videos/v_bG55LSFBA9M.mp4",
"activitynet/videos/v_vbHLA5l_BRc.mkv"
] | 2
|
<answer>2</answer>
|
[
2
] | true
| 1
| 15,805.956434
| 129
|
[
{
"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 sits in a canoe that is docked in shallow water and talks to a camera with the paddle resting across the boat in front of him as a small black dog in a life jacket walks around the boat. The man reaches to pet the dog as the dog walks around and then picks the dog up and holds the dog up by a strap on the top of the life jacket over the water and then sets the dog down back into the boat by the life jacket. The man begins to paddle the boat out into the water as the dog watches on from the side of the boat.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_B0cb0B90Ubg.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_WqDep-4l0yc.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_AZn294ubbps.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_bG55LSFBA9M.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_vbHLA5l_BRc.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: "Three females are dressed in all black in a horizontal line on a red mat. They then begin to move simultaneously together moving their legs and hands with fans in their right hands. All of the girls continue to move stealthy doing numerous lunges and acts of a ninja."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_1B3XsffrM4M.mp4",
"activitynet/videos/v_gUQFX_IydG8.mkv",
"activitynet/videos/v_QMm6gYzsMw8.mp4",
"activitynet/videos/v_hu714U34avg.mp4",
"activitynet/videos/v_K3Z3z8t-RIQ.mp4"
] | 1
|
<answer>1</answer>
|
[
1
] | true
| 1
| 13,752.985604
| 130
|
[
{
"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: \"Three females are dressed in all black in a horizontal line on a red mat. They then begin to move simultaneously together moving their legs and hands with fans in their right hands. All of the girls continue to move stealthy doing numerous lunges and acts of a ninja.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_1B3XsffrM4M.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_gUQFX_IydG8.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_QMm6gYzsMw8.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_hu714U34avg.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_K3Z3z8t-RIQ.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 girl wearing a purple leotard jumps on a balance beam and begins to do her routine. In the middle of it,she does a ballerina spin and almost loses her balance but is able to continue standing up. As she continues,she does more jumps and the judges begin to make marks on their clip boards about her performance. For her last trick,the does three backwards flips and lands on the floor but begins to stumble as she tries to stick her landing."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_oqVNFPUANfs.mp4",
"activitynet/videos/v_5QbiJmDyoM0.mp4",
"activitynet/videos/v_oS4w7-0aH8Q.mp4",
"activitynet/videos/v_ynvCxrj1UNg.mp4",
"activitynet/videos/v_al_NNsjwU-Q.mp4"
] | 4
|
<answer>2</answer>
|
[
2
] | false
| 0
| 20,717.772202
| 131
|
[
{
"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 girl wearing a purple leotard jumps on a balance beam and begins to do her routine. In the middle of it,she does a ballerina spin and almost loses her balance but is able to continue standing up. As she continues,she does more jumps and the judges begin to make marks on their clip boards about her performance. For her last trick,the does three backwards flips and lands on the floor but begins to stumble as she tries to stick her landing.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_oqVNFPUANfs.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_5QbiJmDyoM0.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_oS4w7-0aH8Q.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_ynvCxrj1UNg.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_al_NNsjwU-Q.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 smoking a cigarette while nodding and speaking to the camera. She continues taking puffs off of her cigarette while stroking her hair and occasionally giggling to the camera."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_O0KUnuhLwj0.mp4",
"activitynet/videos/v_v3t4Z5cEgZM.mp4",
"activitynet/videos/v_5SpWmZxECqc.mkv",
"activitynet/videos/v_v2zVnmbPmeQ.mp4",
"activitynet/videos/v_v1Vmf5s42No.mp4"
] | 2
|
<answer>2</answer>
|
[
2
] | true
| 1
| 21,965.989272
| 132
|
[
{
"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 smoking a cigarette while nodding and speaking to the camera. She continues taking puffs off of her cigarette while stroking her hair and occasionally giggling to 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_O0KUnuhLwj0.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_v3t4Z5cEgZM.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_5SpWmZxECqc.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_v2zVnmbPmeQ.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_v1Vmf5s42No.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 nails are seen followed by a shot of brushes and nail polish. A person is then seen wiping polish onto a pad and rubbing the object all over her nails. She then puts a coating over the nail and shows it off again."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_xKePBw5XZHs.mp4",
"activitynet/videos/v_WaVrNbTmbU4.mp4",
"activitynet/videos/v_hL11sP4Hlrg.mp4",
"activitynet/videos/v_JE50XTpCN78.mp4",
"activitynet/videos/v_4VAhZEpQsv8.mp4"
] | 3
|
<answer>3</answer>
|
[
3
] | true
| 1
| 19,084.574957
| 133
|
[
{
"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 nails are seen followed by a shot of brushes and nail polish. A person is then seen wiping polish onto a pad and rubbing the object all over her nails. She then puts a coating over the nail and shows it off again.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_xKePBw5XZHs.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_WaVrNbTmbU4.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_hL11sP4Hlrg.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_JE50XTpCN78.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_4VAhZEpQsv8.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 boy is standing in front of a sink brushing his teeth and dancing. After awhile,he stops to spit in the sink and continues to dance. Finally,the young boy is finished,he spits in the sink again and turns to the camera to smile."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_DinaQYSgbtg.mp4",
"activitynet/videos/v_bfBTnUiGVUo.mp4",
"activitynet/videos/v_rCmpRDbS_O4.mp4",
"activitynet/videos/v_6RePzOd3GvQ.mp4",
"activitynet/videos/v_KjUxjcpIG_Y.mp4"
] | 2
|
<answer>5</answer>
|
[
5
] | false
| 0
| 11,835.685156
| 134
|
[
{
"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 boy is standing in front of a sink brushing his teeth and dancing. After awhile,he stops to spit in the sink and continues to dance. Finally,the young boy is finished,he spits in the sink again and turns to the camera to smile.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_DinaQYSgbtg.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_bfBTnUiGVUo.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_rCmpRDbS_O4.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_6RePzOd3GvQ.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_KjUxjcpIG_Y.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 little girl is wearing pink princess costume while the other girl is wearing wonder woman costume. The girl in pink dress play the hop scotch, pause at the center and continue to hopped back to the start. The girl in pink dress and the girl in wonder woman costume is fixing the tiles, the girl in dress continue to play, then the girl in wonder woman costume play the hop scotch. An adult man appeared and joined the girls and play hop scotch."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_WCCkmuFrSQ0.mp4",
"activitynet/videos/v_CR_79ZjQG_w.mp4",
"activitynet/videos/v_3joaQzU05MY.mp4",
"activitynet/videos/v_VeWdsZb5tog.mp4",
"activitynet/videos/v_z85nM9V4058.mp4"
] | 1
|
<answer>2</answer>
|
[
2
] | false
| 0
| 6,905.899355
| 135
|
[
{
"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 little girl is wearing pink princess costume while the other girl is wearing wonder woman costume. The girl in pink dress play the hop scotch, pause at the center and continue to hopped back to the start. The girl in pink dress and the girl in wonder woman costume is fixing the tiles, the girl in dress continue to play, then the girl in wonder woman costume play the hop scotch. An adult man appeared and joined the girls and play hop scotch.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_WCCkmuFrSQ0.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_CR_79ZjQG_w.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_3joaQzU05MY.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_VeWdsZb5tog.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_z85nM9V4058.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: "There's a large group of people in a band playing their instruments and dancing. There's a small group of cheerleaders on the left who are also cheering. The baton girl begins her routine that includes a lot of twirling and tossing of the baton. The baton girl picks up 2 more batons and continues her routine while using all 3 batons. The baton girl drops 1 of the batons and finishes her routine with 2 batons, waves when she's done and bends down to pick up her batons."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_qA_KTu8oTW4.mp4",
"activitynet/videos/v_6jgWCFWtCfU.mp4",
"activitynet/videos/v_MMVfzKCnpnI.mp4",
"activitynet/videos/v_M0sa3xWhFGo.mp4",
"activitynet/videos/v_Ub88_ql0B78.mkv"
] | 1
|
<answer>5</answer>
|
[
5
] | false
| 0
| 13,600.856205
| 136
|
[
{
"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: \"There's a large group of people in a band playing their instruments and dancing. There's a small group of cheerleaders on the left who are also cheering. The baton girl begins her routine that includes a lot of twirling and tossing of the baton. The baton girl picks up 2 more batons and continues her routine while using all 3 batons. The baton girl drops 1 of the batons and finishes her routine with 2 batons, waves when she's done and bends down to pick up her batons.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_qA_KTu8oTW4.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_6jgWCFWtCfU.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_MMVfzKCnpnI.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_M0sa3xWhFGo.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_Ub88_ql0B78.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 speaking to the camera while holding up a pair of shoes and leads into him untying the shoes and rubbing a rag along it. The man continues to wipe down the shoe using various tools and spray and ends by presenting them."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_tCfu0LplM64.mkv",
"activitynet/videos/v_EFGtb9IDQao.mp4",
"activitynet/videos/v_4OeZViscNp4.mp4",
"activitynet/videos/v_bnQVFmXUx_U.mp4",
"activitynet/videos/v_eRN5gqZFXHI.mp4"
] | 3
|
<answer>1</answer>
|
[
1
] | false
| 0
| 16,688.074578
| 137
|
[
{
"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 speaking to the camera while holding up a pair of shoes and leads into him untying the shoes and rubbing a rag along it. The man continues to wipe down the shoe using various tools and spray and ends by presenting them.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_tCfu0LplM64.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_EFGtb9IDQao.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_4OeZViscNp4.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_bnQVFmXUx_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_eRN5gqZFXHI.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 riding a bike over a trail. They go over several jumps on the trail. A man takes a drink out of a can while sitting on his bike."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_hz3n1wrXYAc.mp4",
"activitynet/videos/v_RHpigjSwhVM.mp4",
"activitynet/videos/v_aYSm25veKTs.mp4",
"activitynet/videos/v_Bh8RcPBQjxo.mp4",
"activitynet/videos/v_1RKExOpIGas.mkv"
] | 2
|
<answer>2</answer>
|
[
2
] | true
| 1
| 21,362.56543
| 138
|
[
{
"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 riding a bike over a trail. They go over several jumps on the trail. A man takes a drink out of a can while sitting on his bike.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_hz3n1wrXYAc.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_RHpigjSwhVM.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_aYSm25veKTs.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_Bh8RcPBQjxo.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_1RKExOpIGas.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 with tattoos is seen speaking to the camera while petting a cat and holding up a pair of nail clippers. The woman then cuts the cat's nails while the cat struggles a bit and holds the cat up in the end."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_En9FemmDusk.mp4",
"activitynet/videos/v_4mzM3JjBJ74.mkv",
"activitynet/videos/v_U36rsW_WhUA.mp4",
"activitynet/videos/v_tfPm9xAZ5z0.mp4",
"activitynet/videos/v_Feo8xSjY5A8.mp4"
] | 2
|
<answer>2</answer>
|
[
2
] | true
| 1
| 19,971.413235
| 139
|
[
{
"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 with tattoos is seen speaking to the camera while petting a cat and holding up a pair of nail clippers. The woman then cuts the cat's nails while the cat struggles a bit and holds the cat up in the end.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_En9FemmDusk.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_4mzM3JjBJ74.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_U36rsW_WhUA.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: "The credits of the clip is shown. A guy moves in front of weights. A group of four pose outside. A man massages a male. A male smells something in a person hands while someone touches his shoulders. Guys lift weights and some celebrate. The audience cheer and applaud. An male individual carries three females. An athlete flexes his guns. A couple dance on stage. Two men pose outside on snowed grass. A guy kisses his thigh."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_Tzm6TEManmQ.mp4",
"activitynet/videos/v_MwQTeFD0OKQ.mp4",
"activitynet/videos/v_OwchMqCYaF4.mp4",
"activitynet/videos/v_rCSFBiXxbVs.mp4",
"activitynet/videos/v_yRgei7gpr-I.mkv"
] | 2
|
<answer>2</answer>
|
[
2
] | true
| 1
| 18,817.480053
| 141
|
[
{
"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 credits of the clip is shown. A guy moves in front of weights. A group of four pose outside. A man massages a male. A male smells something in a person hands while someone touches his shoulders. Guys lift weights and some celebrate. The audience cheer and applaud. An male individual carries three females. An athlete flexes his guns. A couple dance on stage. Two men pose outside on snowed grass. A guy kisses his thigh.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_Tzm6TEManmQ.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_MwQTeFD0OKQ.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_OwchMqCYaF4.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_rCSFBiXxbVs.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_yRgei7gpr-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: "The credits of the clip are shown. A guy flips off a slack line and lands on his body. A man walks on a slack line and falls off. A person does flips around the slack line. A male flips off a brown, picnic table."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_iLaye6q55qk.mp4",
"activitynet/videos/v__D9oML1HvVw.mp4",
"activitynet/videos/v_VR19Scunfhg.mp4",
"activitynet/videos/v_UTiSAR1o2nU.mp4",
"activitynet/videos/v_NzxZdC-63LE.mp4"
] | 2
|
<answer>2</answer>
|
[
2
] | true
| 1
| 26,118.279819
| 140
|
[
{
"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 credits of the clip are shown. A guy flips off a slack line and lands on his body. A man walks on a slack line and falls off. A person does flips around the slack line. A male flips off a brown, picnic table.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_iLaye6q55qk.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__D9oML1HvVw.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_VR19Scunfhg.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_UTiSAR1o2nU.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_NzxZdC-63LE.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 enjoys a soda at a restaurant. The boy taps his bottle of soda. The boy wipes off his mouth with his shirt sleeve."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_KFS_lGlO-Ew.mp4",
"activitynet/videos/v_gUR1wXosHMo.mp4",
"activitynet/videos/v_nNldj5g7W5o.mp4",
"activitynet/videos/v_LCyLWiw7n5Q.mp4",
"activitynet/videos/v_j8lH0saRXl4.mp4"
] | 2
|
<answer>2</answer>
|
[
2
] | true
| 1
| 10,997.481871
| 142
|
[
{
"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 enjoys a soda at a restaurant. The boy taps his bottle of soda. The boy wipes off his mouth with his shirt sleeve.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_KFS_lGlO-Ew.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_gUR1wXosHMo.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_nNldj5g7W5o.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_LCyLWiw7n5Q.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_j8lH0saRXl4.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 man is seen holding two cooking tools in his hand and spinning around food on a pan. The person then serves the food onto a plate and passes it onto a person's hand."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_pwKZRo19Vf0.mp4",
"activitynet/videos/v_ggyGuKFjdxk.mp4",
"activitynet/videos/v_DgcoDX3HbKY.mp4",
"activitynet/videos/v_633ZdPm_GjM.mp4",
"activitynet/videos/v_SrKGO2Xu670.mp4"
] | 3
|
<answer>2</answer>
|
[
2
] | false
| 0
| 17,482.5647
| 144
|
[
{
"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 man is seen holding two cooking tools in his hand and spinning around food on a pan. The person then serves the food onto a plate and passes it onto a person's hand.\"\n\nCandidates:\n[1] 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[2] 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[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_DgcoDX3HbKY.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_633ZdPm_GjM.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_SrKGO2Xu670.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 leafblower is being used to blow leaves across a yard. The mower is moved along slowly, blowing the leaves down a hill."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_bDwGZOk7njI.mp4",
"activitynet/videos/v_GR6Ul2pD8_Y.mp4",
"activitynet/videos/v_2WyRPSKFUi8.mp4",
"activitynet/videos/v_jprf0pE-4uI.mp4",
"activitynet/videos/v_f_BzYUCp4J0.mp4"
] | 5
|
<answer>2</answer>
|
[
2
] | false
| 0
| 18,684.165182
| 145
|
[
{
"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 leafblower is being used to blow leaves across a yard. The mower is moved along slowly, blowing the leaves down a hill.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_bDwGZOk7njI.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_GR6Ul2pD8_Y.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_2WyRPSKFUi8.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_jprf0pE-4uI.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_f_BzYUCp4J0.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 stands in a room. She is wearing a blue polka dot dress. She is playing a violin. She smiles at the end."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_1hB5jVAhSDE.mp4",
"activitynet/videos/v_z-ttrQ38mOc.mp4",
"activitynet/videos/v_OPp3DqFq0O0.mp4",
"activitynet/videos/v_r1mrueEHDDE.mp4",
"activitynet/videos/v_-E9YQ_Uhu50.mp4"
] | 3
|
<answer>2</answer>
|
[
2
] | false
| 0
| 28,549.432118
| 143
|
[
{
"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 stands in a room. She is wearing a blue polka dot dress. She is playing a violin. She smiles at the end.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_1hB5jVAhSDE.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_z-ttrQ38mOc.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_OPp3DqFq0O0.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_r1mrueEHDDE.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_-E9YQ_Uhu50.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 wearing black shirt is in a courtroom playing raketball. woman wearing a yellow shirt is on a courtroom playing raketball. two people are standing on a small court holding tennis rackets and playing raketball."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_R7DhZaY3A08.mp4",
"activitynet/videos/v__VPf75tGIHQ.mp4",
"activitynet/videos/v_rZGxJN2AOQY.mp4",
"activitynet/videos/v_4At1Vd-0lWE.mp4",
"activitynet/videos/v_IiCN1md2MV4.mp4"
] | 1
|
<answer>4</answer>
|
[
4
] | false
| 0
| 9,239.899377
| 147
|
[
{
"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 wearing black shirt is in a courtroom playing raketball. woman wearing a yellow shirt is on a courtroom playing raketball. two people are standing on a small court holding tennis rackets and playing raketball.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_R7DhZaY3A08.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__VPf75tGIHQ.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_rZGxJN2AOQY.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_4At1Vd-0lWE.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_IiCN1md2MV4.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 talks holding an empty cup, and coffee fills in cups. The woman continues talking while measures off coffee display the time in time in the screen. Then, appears a picture of a cup of coffee writing "red eye" , after the woman continues talking."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_Pv6oIFroaCQ.mp4",
"activitynet/videos/v_hmb86jpgWfE.mp4",
"activitynet/videos/v_R-RQx5pbMvo.mp4",
"activitynet/videos/v_IZx-EMbylmM.mp4",
"activitynet/videos/v_aa5jHg4E3O0.mp4"
] | 2
|
<answer>4</answer>
|
[
4
] | false
| 0
| 16,987.90399
| 146
|
[
{
"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 talks holding an empty cup, and coffee fills in cups. The woman continues talking while measures off coffee display the time in time in the screen. Then, appears a picture of a cup of coffee writing \"red eye\" , after the woman 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_Pv6oIFroaCQ.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_hmb86jpgWfE.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_R-RQx5pbMvo.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_IZx-EMbylmM.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_aa5jHg4E3O0.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 stands in front of a canvas holding a paint brush. He dips the paint brush in a bucket and wipes it off on a towel. He then continues to clean the brush."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_sEGceBU8icE.mp4",
"activitynet/videos/v_4zYY4abpCgI.mp4",
"activitynet/videos/v_ebmi7XJA8Oo.mkv",
"activitynet/videos/v_WzDnorAzWVU.mp4",
"activitynet/videos/v_ev-RTtbVjFI.mp4"
] | 4
|
<answer>4</answer>
|
[
4
] | true
| 1
| 14,592.606722
| 148
|
[
{
"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 stands in front of a canvas holding a paint brush. He dips the paint brush in a bucket and wipes it off on a towel. He then continues to clean the brush.\"\n\nCandidates:\n[1] 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[2] 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[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_ebmi7XJA8Oo.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_WzDnorAzWVU.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 man is standing in the middle of the road, wearing stilts. He walks around, showing how they work. He then runs down the street, bouncing and springing on the stilts. He stops and speaks to the camera, smiling."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_LqCg09IRp-o.mp4",
"activitynet/videos/v_xXTfM9xXFQM.mp4",
"activitynet/videos/v_M6FdaEstXbI.mp4",
"activitynet/videos/v_wuknZBoyMRE.mp4",
"activitynet/videos/v_xDD9rWISPpk.mp4"
] | 2
|
<answer>2</answer>
|
[
2
] | true
| 1
| 15,331.090988
| 149
|
[
{
"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 in the middle of the road, wearing stilts. He walks around, showing how they work. He then runs down the street, bouncing and springing on the stilts. He stops and speaks to the camera, smiling.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_LqCg09IRp-o.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_xXTfM9xXFQM.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_wuknZBoyMRE.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_xDD9rWISPpk.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 making a circle play pipes in an event ma. Persons stand next a drum and other instruments. A woman takes pictures of people playing pipes."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_UR1e1MIRvvc.mp4",
"activitynet/videos/v_Ti1ZaH0VGfg.mp4",
"activitynet/videos/v_TVmuh_sR1KI.mp4",
"activitynet/videos/v_tXLvsYsWCoY.mp4",
"activitynet/videos/v_MSSb3wPd5hM.mp4"
] | 2
|
<answer>2</answer>
|
[
2
] | true
| 1
| 17,205.545372
| 150
|
[
{
"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 making a circle play pipes in an event ma. Persons stand next a drum and other instruments. A woman takes pictures of people playing pipes.\"\n\nCandidates:\n[1] 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[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_Ti1ZaH0VGfg.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_TVmuh_sR1KI.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_tXLvsYsWCoY.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_MSSb3wPd5hM.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 machine is shown followed by a man speaking to the camera. The man uses the tool around a living room while also cleaning out a vent. He uses the tool more while speaking to the camera as well as taking it apart to show how it works."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_nkrA8sJydF0.mp4",
"activitynet/videos/v_ep2Kyk8CHT8.mp4",
"activitynet/videos/v_50b9lVikSeo.mp4",
"activitynet/videos/v_-0r0HEwAYiQ.mp4",
"activitynet/videos/v_DfFqlrv7F2g.mp4"
] | 1
|
<answer>2</answer>
|
[
2
] | false
| 0
| 15,103.29596
| 151
|
[
{
"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 machine is shown followed by a man speaking to the camera. The man uses the tool around a living room while also cleaning out a vent. He uses the tool more while speaking to the camera as well as taking it apart to show how it works.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_nkrA8sJydF0.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_ep2Kyk8CHT8.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_50b9lVikSeo.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_-0r0HEwAYiQ.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_DfFqlrv7F2g.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 people are seen swimming around a pool when a ball comes into frame. The group of people then throw a ball around the area while a person walks in and out of frame. The people continue to play with one another while throwing the ball back and fourth."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_v6go4RA0ZB4.mp4",
"activitynet/videos/v_DFJBJkCR0Bk.mp4",
"activitynet/videos/v_qKG1mU0Feug.mp4",
"activitynet/videos/v_m5NK0eErs90.mp4",
"activitynet/videos/v_OHNH7IV0768.mp4"
] | 2
|
<answer>2</answer>
|
[
2
] | true
| 1
| 13,581.877497
| 152
|
[
{
"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 people are seen swimming around a pool when a ball comes into frame. The group of people then throw a ball around the area while a person walks in and out of frame. The people continue to play with one another while throwing the ball back and fourth.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_v6go4RA0ZB4.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_DFJBJkCR0Bk.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_qKG1mU0Feug.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_m5NK0eErs90.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_OHNH7IV0768.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 bend to raise a weight until his shoulders to rest, and then raise the weight above his head. After, the boy left fall the weight to the floor. A man enters the gym."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_jzCnWUUUviE.mp4",
"activitynet/videos/v_SKtUq_1cOSs.mp4",
"activitynet/videos/v_gIhVeU8xbrs.mp4",
"activitynet/videos/v_U9Dcet1qdRE.mp4",
"activitynet/videos/v_E7C91KoML-o.mp4"
] | 1
|
<answer>1</answer>
|
[
1
] | true
| 1
| 9,083.441728
| 153
|
[
{
"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 bend to raise a weight until his shoulders to rest, and then raise the weight above his head. After, the boy left fall the weight to the floor. A man enters the gym.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_jzCnWUUUviE.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_SKtUq_1cOSs.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_gIhVeU8xbrs.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_U9Dcet1qdRE.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_E7C91KoML-o.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 chef's hat is seen speaking to the camera and leads into a completed cake made and various ingredients being poured into a bowl. The man blends all the ingredients together and lays the dough out on a pan and adding more chocolate chips. A shot of the giant cookie baked is shown and leads into the man holding it in front of the camera and eating it."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_v6Ui5kgi2OI.mp4",
"activitynet/videos/v_qx1FNJxiUuE.mp4",
"activitynet/videos/v_WPVb8fYLFUM.mp4",
"activitynet/videos/v_v7OluHKECRA.mp4",
"activitynet/videos/v_J2gJYNO2qh8.mp4"
] | 3
|
<answer>3</answer>
|
[
3
] | true
| 1
| 23,478.222479
| 154
|
[
{
"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 chef's hat is seen speaking to the camera and leads into a completed cake made and various ingredients being poured into a bowl. The man blends all the ingredients together and lays the dough out on a pan and adding more chocolate chips. A shot of the giant cookie baked is shown and leads into the man holding it in front of the camera and eating it.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_v6Ui5kgi2OI.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_qx1FNJxiUuE.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_WPVb8fYLFUM.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_v7OluHKECRA.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 is seen pushing a large machine around a wooded area and cutting the grass along the trees. A video shows how the machine works and leads into a man speaking to the camera and shots of the machine working."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_vigHVj40dO4.mkv",
"activitynet/videos/v_f_BzYUCp4J0.mp4",
"activitynet/videos/v_ur873jaQO3A.mkv",
"activitynet/videos/v_2WyRPSKFUi8.mp4",
"activitynet/videos/v_CJWSed5v4jE.mkv"
] | 1
|
<answer>1</answer>
|
[
1
] | true
| 1
| 23,320.275198
| 156
|
[
{
"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 pushing a large machine around a wooded area and cutting the grass along the trees. A video shows how the machine works and leads into a man speaking to the camera and shots of the machine working.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_vigHVj40dO4.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_f_BzYUCp4J0.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_ur873jaQO3A.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_2WyRPSKFUi8.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_CJWSed5v4jE.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 sitting in a field and touching a pole in the ground. The man stands up and touches the vines. The man cuts a few parts of the vine. The man waves his hands over the vine. The man stops and talks to the camera. The screen fades to black."
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_5O9myPtSriI.mp4",
"activitynet/videos/v_i0Z8I2WCLNk.mp4",
"activitynet/videos/v_ddLFSNa3ci0.mp4",
"activitynet/videos/v_UxR9fdD0Vzw.mp4"
] | 3
|
<answer>3</answer>
|
[
3
] | true
| 1
| 30,065.844936
| 155
|
[
{
"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 sitting in a field and touching a pole in the ground. The man stands up and touches the vines. The man cuts a few parts of the vine. The man waves his hands over the vine. The man stops and talks to the camera. The screen fades to black.\"\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_5O9myPtSriI.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_i0Z8I2WCLNk.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_ddLFSNa3ci0.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_UxR9fdD0Vzw.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 shave his right leg with a shaver machine. Then the boy raises the pants of his left leg and shows both legs while talking. A person enters in the bedroom."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_gh0GD6OvLHE.mp4",
"activitynet/videos/v_-MFzpFMdWZs.mp4",
"activitynet/videos/v_0fsMeZoZzJI.mp4",
"activitynet/videos/v_-U4lNtzVQ8s.mp4",
"activitynet/videos/v_NNOsdZr802w.mp4"
] | 2
|
<answer>5</answer>
|
[
5
] | false
| 0
| 8,334.354134
| 158
|
[
{
"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 shave his right leg with a shaver machine. Then the boy raises the pants of his left leg and shows both legs while talking. A person enters in the bedroom.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_gh0GD6OvLHE.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_-MFzpFMdWZs.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_0fsMeZoZzJI.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_-U4lNtzVQ8s.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_NNOsdZr802w.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 are tearing up the carpet in a store of some kind then move the carpet out and paint the floor gray. The men then add gray carpet to the floor and a man vacuums. We switch and see them rip up and lay new carpet from another angle. We see a man vacuuming the new carpet."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_9hE6VRD3qXQ.mp4",
"activitynet/videos/v__ye90Ou8SnE.mp4",
"activitynet/videos/v_EXr5QXCpkYY.mp4",
"activitynet/videos/v_doZb3RlLSts.mp4",
"activitynet/videos/v_dwCeFVAaP9c.mp4"
] | 3
|
<answer>3</answer>
|
[
3
] | true
| 1
| 12,052.181581
| 157
|
[
{
"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 are tearing up the carpet in a store of some kind then move the carpet out and paint the floor gray. The men then add gray carpet to the floor and a man vacuums. We switch and see them rip up and lay new carpet from another angle. We see a man vacuuming the new carpet.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_9hE6VRD3qXQ.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__ye90Ou8SnE.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_EXr5QXCpkYY.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_doZb3RlLSts.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_dwCeFVAaP9c.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 drive bumper cars around a bumper car rink in an amusement park. A yellow bumper car hits another bumper car. A red bumper car hits the outer wall and stalls out."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v__crwKCjKRjg.mp4",
"activitynet/videos/v_yQ2AirKmnTM.mkv",
"activitynet/videos/v_ZjvmWr5LoFw.mp4",
"activitynet/videos/v_HCphw9_Jku8.mp4",
"activitynet/videos/v_nX-GvQmf5Tc.mp4"
] | 2
|
<answer>2</answer>
|
[
2
] | true
| 1
| 9,846.846463
| 159
|
[
{
"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 drive bumper cars around a bumper car rink in an amusement park. A yellow bumper car hits another bumper car. A red bumper car hits the outer wall and stalls out.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v__crwKCjKRjg.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_yQ2AirKmnTM.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_ZjvmWr5LoFw.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_HCphw9_Jku8.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_nX-GvQmf5Tc.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's ear is shown close up. A man shows the tool he is going to use to pierce his ear. He inserts the needle, piercing his ear. He then adds a tube, creating a gauge."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v__KOVk8iGbrA.mp4",
"activitynet/videos/v_DgJ-GG1Agyw.mp4",
"activitynet/videos/v_juIOpLYnW64.mp4",
"activitynet/videos/v_Qre7RVxEn78.mp4",
"activitynet/videos/v_MbEtgOmOY-4.mp4"
] | 4
|
<answer>2</answer>
|
[
2
] | false
| 0
| 13,023.819635
| 160
|
[
{
"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's ear is shown close up. A man shows the tool he is going to use to pierce his ear. He inserts the needle, piercing his ear. He then adds a tube, creating a gauge.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v__KOVk8iGbrA.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_DgJ-GG1Agyw.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_juIOpLYnW64.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_Qre7RVxEn78.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_MbEtgOmOY-4.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: "Bmx riders are at the gate waiting for it to be brought down to start racing. Once the gates are down they are on the move riding through the course. Group after group starts racing and moving along really fast. The peddle as fast as they can trying to stay ahead of one another."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_x8AR0FD5Jqo.mp4",
"activitynet/videos/v_b380n1dci9I.mp4",
"activitynet/videos/v_r015El3onHw.mp4",
"activitynet/videos/v_my9Z4bun_Dk.mp4",
"activitynet/videos/v_mFDC1CLt6B4.mp4"
] | 1
|
<answer>1</answer>
|
[
1
] | true
| 1
| 19,023.582668
| 161
|
[
{
"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: \"Bmx riders are at the gate waiting for it to be brought down to start racing. Once the gates are down they are on the move riding through the course. Group after group starts racing and moving along really fast. The peddle as fast as they can trying to stay ahead of 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_x8AR0FD5Jqo.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_b380n1dci9I.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_r015El3onHw.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_my9Z4bun_Dk.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_mFDC1CLt6B4.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 sax playing is on a stage playing. There is a close up of the hands. He is dancing as he plays the sax. There is a band behind him. The camera does an over head shot. The player is finished playing."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_s0N0PzdwVik.mp4",
"activitynet/videos/v_clSku91LoQQ.mp4",
"activitynet/videos/v_0EdDWY0Zuqw.mp4",
"activitynet/videos/v_aN9vCyXMbb8.mp4",
"activitynet/videos/v_Zcq_xLi2NGo.mp4"
] | 5
|
<answer>3</answer>
|
[
3
] | false
| 0
| 17,385.184224
| 163
|
[
{
"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 sax playing is on a stage playing. There is a close up of the hands. He is dancing as he plays the sax. There is a band behind him. The camera does an over head shot. The player is finished playing.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_s0N0PzdwVik.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_clSku91LoQQ.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_0EdDWY0Zuqw.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_aN9vCyXMbb8.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_Zcq_xLi2NGo.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 swirl appears on a black title screen A person is riding a kite surf in the sea. We see the man perform a flip with a title screen. We see the man ride the kitesurf and perform the same flip. The man performs another flip, the video rewinds, and we see the flip again. The ending video credits load."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_g-rw2Kyh9xo.mkv",
"activitynet/videos/v_Y0G_wA38HkI.mp4",
"activitynet/videos/v_plE3KNmuwj4.mp4",
"activitynet/videos/v_TyHLBe6__rc.mp4",
"activitynet/videos/v_hPJw9_nPo_s.mkv"
] | 2
|
<answer>4</answer>
|
[
4
] | false
| 0
| 25,047.035379
| 162
|
[
{
"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 swirl appears on a black title screen A person is riding a kite surf in the sea. We see the man perform a flip with a title screen. We see the man ride the kitesurf and perform the same flip. The man performs another flip, the video rewinds, and we see the flip again. The ending video credits load.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_g-rw2Kyh9xo.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_Y0G_wA38HkI.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_plE3KNmuwj4.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_TyHLBe6__rc.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_hPJw9_nPo_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 woman with long hair is seen poking at her cheeks and taking a puff out of a hookah. She then blows smoke rings into the camera and looks down smiling."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_t1-GV2bAL4I.mp4",
"activitynet/videos/v_9Bo7Hr77DgA.mp4",
"activitynet/videos/v_jwndE_xn8sA.mp4",
"activitynet/videos/v_v3t4Z5cEgZM.mp4",
"activitynet/videos/v_4OfhHE72V8c.mp4"
] | 3
|
<answer>3</answer>
|
[
3
] | true
| 1
| 13,034.696088
| 165
|
[
{
"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 woman with long hair is seen poking at her cheeks and taking a puff out of a hookah. She then blows smoke rings into the camera and looks down smiling.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_t1-GV2bAL4I.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_9Bo7Hr77DgA.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_jwndE_xn8sA.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_v3t4Z5cEgZM.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_4OfhHE72V8c.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 snowy mountain top. They are skiing down the numerous hills together. We see them flip and turn sharply in the driven snow."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_slHv7r8A4OI.mp4",
"activitynet/videos/v_iAIl5eawd6I.mp4",
"activitynet/videos/v_3nrianTc060.mp4",
"activitynet/videos/v_cY3QbnSeu9k.mp4",
"activitynet/videos/v_6VygM9-XgAk.mp4"
] | 1
|
<answer>1</answer>
|
[
1
] | true
| 1
| 24,727.059741
| 164
|
[
{
"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 snowy mountain top. They are skiing down the numerous hills together. We see them flip and turn sharply in the driven snow.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_slHv7r8A4OI.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_iAIl5eawd6I.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_3nrianTc060.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_cY3QbnSeu9k.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_6VygM9-XgAk.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 of a house installing new carpet on the floor. He is using a measuring stick and some plier to cut off a piece of metal which he then hammers down with nails underneath the carpet. He rolls in the end of the carpet inside of the metal so it does not show. Then he gets the hammer again and he hammers down the carpet so that it sticks in place."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_xf9iLflgRro.mkv",
"activitynet/videos/v_vL8s-b4eJiU.mp4",
"activitynet/videos/v_iUVz4A5oblQ.mp4",
"activitynet/videos/v_dZyb8t-4ATQ.mp4",
"activitynet/videos/v_O-upcCp0jIs.mp4"
] | 5
|
<answer>3</answer>
|
[
3
] | false
| 0
| 17,756.300657
| 166
|
[
{
"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 of a house installing new carpet on the floor. He is using a measuring stick and some plier to cut off a piece of metal which he then hammers down with nails underneath the carpet. He rolls in the end of the carpet inside of the metal so it does not show. Then he gets the hammer again and he hammers down the carpet so that it sticks in place.\"\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_vL8s-b4eJiU.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_iUVz4A5oblQ.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_dZyb8t-4ATQ.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_O-upcCp0jIs.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 bucket is seen sitting in a yard followed by a person moving up and down the yard. The person uses a tool all along the yard while the camera captures their movements. They pull out another object after a day and continue moving around the yard."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_1hiyhNqakMI.mp4",
"activitynet/videos/v_SBj7yuFEwQI.mp4",
"activitynet/videos/v_JSqJmZPqDy8.mp4",
"activitynet/videos/v_2Sev8z4P7pE.mp4",
"activitynet/videos/v_Q_32kySHzCQ.mp4"
] | 1
|
<answer>3</answer>
|
[
3
] | false
| 0
| 10,343.66887
| 167
|
[
{
"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 bucket is seen sitting in a yard followed by a person moving up and down the yard. The person uses a tool all along the yard while the camera captures their movements. They pull out another object after a day and continue moving around 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_1hiyhNqakMI.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_SBj7yuFEwQI.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_JSqJmZPqDy8.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_Q_32kySHzCQ.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 speaking to the camera while the man pans down to their feet and one throws darts. The man continuously throws darts while the other speaks to the camera."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_EQMDnhIKU4w.mkv",
"activitynet/videos/v_5ptxyeHlcwM.mp4",
"activitynet/videos/v_1PQiq8zajCE.mp4",
"activitynet/videos/v_9XmzbuByY_E.mp4",
"activitynet/videos/v_lVu-4SKcb4c.mp4"
] | 1
|
<answer>3</answer>
|
[
3
] | false
| 0
| 10,135.276144
| 168
|
[
{
"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 speaking to the camera while the man pans down to their feet and one throws darts. The man continuously throws darts while the other speaks to 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_EQMDnhIKU4w.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_5ptxyeHlcwM.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_1PQiq8zajCE.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_9XmzbuByY_E.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_lVu-4SKcb4c.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: "At a skating park a older gentlemen talks about skating while he is outside. He gets on his board and starts to skate all around the park looking pretty hip. He talks some more and he gets done to do some maintenance on his board or something. He seems like a young soul trying to stay fit enough to skate."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_lM4FQ_FqEhQ.mp4",
"activitynet/videos/v_txsupdxCToQ.mp4",
"activitynet/videos/v_6fuOwhx91zM.mp4",
"activitynet/videos/v_pXcFBfv5Sf4.mp4",
"activitynet/videos/v_oQuAwR_t5Ig.mp4"
] | 1
|
<answer>1</answer>
|
[
1
] | true
| 1
| 18,685.170327
| 169
|
[
{
"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: \"At a skating park a older gentlemen talks about skating while he is outside. He gets on his board and starts to skate all around the park looking pretty hip. He talks some more and he gets done to do some maintenance on his board or something. He seems like a young soul trying to stay fit enough to skate.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_lM4FQ_FqEhQ.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_txsupdxCToQ.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_6fuOwhx91zM.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_pXcFBfv5Sf4.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_oQuAwR_t5Ig.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 gets her hair brushed combed brushed and styled by an unseen person's hands as a camera catches the process from difference angles. A woman, has curlers placed in her hair by a pair of hands holding a comb. A pair of hands now brushes the woman's hair to the side and also brushes the ends of the hair. The woman's finished hairdo is captured by the camera in still shots at many angles."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_sMITf5WBIxM.mp4",
"activitynet/videos/v_LRhkbJ9dcP8.mp4",
"activitynet/videos/v_oKauZV0DHHk.mp4",
"activitynet/videos/v_2oc0OBWkYfg.mp4",
"activitynet/videos/v_f1kY1-9XR1k.mp4"
] | 4
|
<answer>2</answer>
|
[
2
] | false
| 0
| 17,798.233947
| 171
|
[
{
"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 gets her hair brushed combed brushed and styled by an unseen person's hands as a camera catches the process from difference angles. A woman, has curlers placed in her hair by a pair of hands holding a comb. A pair of hands now brushes the woman's hair to the side and also brushes the ends of the hair. The woman's finished hairdo is captured by the camera in still shots at many angles.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_sMITf5WBIxM.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_LRhkbJ9dcP8.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_oKauZV0DHHk.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_2oc0OBWkYfg.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_f1kY1-9XR1k.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 floats down a river. His feet are propped up. Several trees line the river. Two other people are shown floating as well."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_yJSQmNSFlNI.mkv",
"activitynet/videos/v_akUXL2VzFEs.mp4",
"activitynet/videos/v_1buoiCgXG1Q.mp4",
"activitynet/videos/v_DeoqC3oVV38.mkv",
"activitynet/videos/v_9GwsrWUq7mY.mp4"
] | 2
|
<answer>4</answer>
|
[
4
] | false
| 0
| 25,433.32196
| 170
|
[
{
"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 floats down a river. His feet are propped up. Several trees line the river. Two other people are shown floating as well.\"\n\nCandidates:\n[1] 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[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_akUXL2VzFEs.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_1buoiCgXG1Q.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_DeoqC3oVV38.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_9GwsrWUq7mY.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 credits of the clip are shown. People are floating on inflated tires. A guy swims across the body of water. A lady puts her hand in front of the camera. People are standing in a shallow part of the body of water, and a guy falls into the water. Three people pose on a rock. People take the inflated tires out of the water. A woman splashes water on a man floating. People use a rope to dive into the water."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_I8nK8c7k9ko.mp4",
"activitynet/videos/v_py26bxAfOEg.mp4",
"activitynet/videos/v_sG3JpMuXFnU.mp4",
"activitynet/videos/v_V4srMOGRlU8.mp4",
"activitynet/videos/v_HfN967uah8o.mp4"
] | 2
|
<answer>2</answer>
|
[
2
] | true
| 1
| 21,738.000359
| 172
|
[
{
"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 credits of the clip are shown. People are floating on inflated tires. A guy swims across the body of water. A lady puts her hand in front of the camera. People are standing in a shallow part of the body of water, and a guy falls into the water. Three people pose on a rock. People take the inflated tires out of the water. A woman splashes water on a man floating. People use a rope to dive into the water.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_I8nK8c7k9ko.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_py26bxAfOEg.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_sG3JpMuXFnU.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_V4srMOGRlU8.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_HfN967uah8o.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 dribbling a basketball around orange cones. He shoots and makes a basket several times."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_bOBQLGfEeyg.mp4",
"activitynet/videos/v_D4wcmmQsPng.mp4",
"activitynet/videos/v_1926p23ooUM.mp4",
"activitynet/videos/v_L21zcZlFfIY.mp4",
"activitynet/videos/v_ZlwkO1oFBHw.mp4"
] | 3
|
<answer>3</answer>
|
[
3
] | true
| 1
| 18,034.195351
| 173
|
[
{
"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 dribbling a basketball around orange cones. He shoots and makes a basket several times.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_bOBQLGfEeyg.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_D4wcmmQsPng.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_1926p23ooUM.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_L21zcZlFfIY.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_ZlwkO1oFBHw.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 young children are standing in line indoors in what looks like a living room. The little girl is standing closest to the hopscotch mat and she throws her toy onto the mat and then begins jumping until she meets the end of the mat then turns around and heads back to the point she started and her turn is over. The little boy goes next and he throws the toy onto the mat and begins jumping to the end of the mat, then turns around and jumps back towards his starting point. The little girl steps in front of the boy and gets into motion to start another turn on the hop scotch mat."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_lKKimizxQJM.mp4",
"activitynet/videos/v_YozbZM_nA0c.mp4",
"activitynet/videos/v_OvGxDaayPcw.mp4",
"activitynet/videos/v_3joaQzU05MY.mp4",
"activitynet/videos/v_uBmUiouilQY.mp4"
] | 3
|
<answer>3</answer>
|
[
3
] | true
| 1
| 14,871.340793
| 174
|
[
{
"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 young children are standing in line indoors in what looks like a living room. The little girl is standing closest to the hopscotch mat and she throws her toy onto the mat and then begins jumping until she meets the end of the mat then turns around and heads back to the point she started and her turn is over. The little boy goes next and he throws the toy onto the mat and begins jumping to the end of the mat, then turns around and jumps back towards his starting point. The little girl steps in front of the boy and gets into motion to start another turn on the hop scotch mat.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_lKKimizxQJM.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_YozbZM_nA0c.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_OvGxDaayPcw.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_3joaQzU05MY.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_uBmUiouilQY.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 and child are sliding down a slide. The both raise their hands in the air. When they reach the end, the child stands up."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_5I0K3y27EUM.mp4",
"activitynet/videos/v_bc-DycGxV9E.mp4",
"activitynet/videos/v_q2KR3lzTrq4.mkv",
"activitynet/videos/v_dMjOeGJBF9M.mp4",
"activitynet/videos/v_Zc44Ddk2NG8.mp4"
] | 1
|
<answer>2</answer>
|
[
2
] | false
| 0
| 7,250.93127
| 175
|
[
{
"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 and child are sliding down a slide. The both raise their hands in the air. When they reach the end, the child stands up.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_5I0K3y27EUM.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_bc-DycGxV9E.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_q2KR3lzTrq4.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_dMjOeGJBF9M.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_Zc44Ddk2NG8.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 child is seen sitting before a set of drums and smiling into the camera. She puts her hands up to her mouth and continues playing while pausing every few moments to speak to the camera."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_4ACqWG_p1bI.mp4",
"activitynet/videos/v_iY7bZQnHXlk.mp4",
"activitynet/videos/v_yISeNkFiVAg.mp4",
"activitynet/videos/v_bb-DPA34qvw.mp4",
"activitynet/videos/v_9q6wWG6ql4E.mp4"
] | 4
|
<answer>4</answer>
|
[
4
] | true
| 1
| 19,563.324835
| 176
|
[
{
"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 child is seen sitting before a set of drums and smiling into the camera. She puts her hands up to her mouth and continues playing while pausing every few moments to speak to 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_4ACqWG_p1bI.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_iY7bZQnHXlk.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_yISeNkFiVAg.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_bb-DPA34qvw.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_9q6wWG6ql4E.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 wrestlers are seen holding one another in a ring while a ref stands behind them. One man picks up the other and throws him onto the ground."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_K98WGaMR4eM.mkv",
"activitynet/videos/v_hltWAq_Odxk.mp4",
"activitynet/videos/v_r8hXEpP7HH0.mp4",
"activitynet/videos/v_MjljlkQaHh4.mp4",
"activitynet/videos/v_8nyOw9vBh2E.mp4"
] | 2
|
<answer>2</answer>
|
[
2
] | true
| 1
| 13,519.780056
| 178
|
[
{
"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 wrestlers are seen holding one another in a ring while a ref stands behind them. One man picks up the other and throws him onto 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_K98WGaMR4eM.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_hltWAq_Odxk.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_r8hXEpP7HH0.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_MjljlkQaHh4.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_8nyOw9vBh2E.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 mountain, then a man talks and then put on shoes. The man claims a mountain only using his fingers, and without carry any safety materials he reaches the top of the mountain. Then, the man drives a car and arrives to a mountain. After, the man climbs a mountain carrying a backpack, he arrives a plateau and prepares food to eat. Next, the man continues climbing the steep mountain until the top of the mountain."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_VXCV3KUtCdk.mp4",
"activitynet/videos/v_6Yn2U58qxPs.mp4",
"activitynet/videos/v_ZGL-PmMopeM.mp4",
"activitynet/videos/v_zOGg5-Mll4o.mp4",
"activitynet/videos/v_lfH_S2LTEXA.mp4"
] | 1
|
<answer>2</answer>
|
[
2
] | false
| 0
| 26,089.827926
| 177
|
[
{
"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 mountain, then a man talks and then put on shoes. The man claims a mountain only using his fingers, and without carry any safety materials he reaches the top of the mountain. Then, the man drives a car and arrives to a mountain. After, the man climbs a mountain carrying a backpack, he arrives a plateau and prepares food to eat. Next, the man continues climbing the steep mountain until the top of the mountain.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_VXCV3KUtCdk.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_6Yn2U58qxPs.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_ZGL-PmMopeM.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_lfH_S2LTEXA.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 picture of a red arrow is shown. A picture of a person kneeling down on a roof holding a nail gun."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_4CsTbXdERSU.mp4",
"activitynet/videos/v_UWTpfygMUQw.mp4",
"activitynet/videos/v_VUvEWwghANE.mp4",
"activitynet/videos/v_sqcJOpPrexQ.mp4",
"activitynet/videos/v_6rOmYOU7748.mp4"
] | 5
|
<answer>3</answer>
|
[
3
] | false
| 0
| 15,303.039607
| 179
|
[
{
"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 picture of a red arrow is shown. A picture of a person kneeling down on a roof holding a nail gun.\"\n\nCandidates:\n[1] 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
}
},
{
"type": "text",
"text": "\n[2] 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[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_VUvEWwghANE.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_sqcJOpPrexQ.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_6rOmYOU7748.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 sitting on a bike in a room. The man quickly removes the front tire, fixes it and puts it back on the bike. The man airs up the tire, grabs his tools and puts the cap on. He stops and gets off the bike. The people watching him all start to clap. The man bows, lifts his glasses and starts to walk away."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_pIk9qMEyEd4.mp4",
"activitynet/videos/v_xIhTY02lRSE.mp4",
"activitynet/videos/v_KNyM0KvDHMM.mp4",
"activitynet/videos/v_kShrO0yutUQ.mp4",
"activitynet/videos/v_A7ER02-zr54.mp4"
] | 2
|
<answer>2</answer>
|
[
2
] | true
| 1
| 23,231.599613
| 180
|
[
{
"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 sitting on a bike in a room. The man quickly removes the front tire, fixes it and puts it back on the bike. The man airs up the tire, grabs his tools and puts the cap on. He stops and gets off the bike. The people watching him all start to clap. The man bows, lifts his glasses and starts to walk away.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_pIk9qMEyEd4.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_xIhTY02lRSE.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_KNyM0KvDHMM.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_kShrO0yutUQ.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_A7ER02-zr54.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 girl wearing a blue ribbon in her hair standing in front of a rock garden demonstrates a gymnastic move called a Power Hurdle. The young gymnast demonstrates an exercise called a back-bend move. The young gymnast with the blue ribbon in her hair performs a back walkover and a front walkover and a forward roll."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_iAPv-QSvZF4.mp4",
"activitynet/videos/v_1L_4N307nBk.mp4",
"activitynet/videos/v_oS4w7-0aH8Q.mp4",
"activitynet/videos/v_C10_qXWxpsk.mp4",
"activitynet/videos/v_7rpq2RXAoKE.mp4"
] | 4
|
<answer>4</answer>
|
[
4
] | true
| 1
| 16,836.80544
| 181
|
[
{
"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 girl wearing a blue ribbon in her hair standing in front of a rock garden demonstrates a gymnastic move called a Power Hurdle. The young gymnast demonstrates an exercise called a back-bend move. The young gymnast with the blue ribbon in her hair performs a back walkover and a front walkover and a forward roll.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_iAPv-QSvZF4.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_1L_4N307nBk.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_oS4w7-0aH8Q.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_C10_qXWxpsk.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_7rpq2RXAoKE.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 intro leads into a basketball player holding a ball and then running up and shooting a basket. The man continues running back and fourth while shooting the baskets and ends with text across the screen."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_ZlwkO1oFBHw.mp4",
"activitynet/videos/v_1926p23ooUM.mp4",
"activitynet/videos/v_q0KrlywYHM8.mkv",
"activitynet/videos/v_agM7yjqVKo0.mp4",
"activitynet/videos/v_0GWJ-VHFlTk.mp4"
] | 5
|
<answer>4</answer>
|
[
4
] | false
| 0
| 15,975.183923
| 182
|
[
{
"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 intro leads into a basketball player holding a ball and then running up and shooting a basket. The man continues running back and fourth while shooting the baskets and ends with text across 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_ZlwkO1oFBHw.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_1926p23ooUM.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_q0KrlywYHM8.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_agM7yjqVKo0.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_0GWJ-VHFlTk.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 kid in a yellow shirt is skating around a hockey rink. He is holding a hockey stick. He shoots at the net and scores."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_D4wcmmQsPng.mp4",
"activitynet/videos/v_8OEts-YLeW0.mkv",
"activitynet/videos/v_Wy0u0amd4Ko.mp4",
"activitynet/videos/v_BMxtjh9E7BY.mp4",
"activitynet/videos/v_yu1XjQUctiM.mp4"
] | 1
|
<answer>3</answer>
|
[
3
] | false
| 0
| 14,205.496098
| 183
|
[
{
"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 kid in a yellow shirt is skating around a hockey rink. He is holding a hockey stick. He shoots at the net and scores.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_D4wcmmQsPng.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_8OEts-YLeW0.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_Wy0u0amd4Ko.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_BMxtjh9E7BY.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_yu1XjQUctiM.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 sitting outside on the bleachers watching a game of water polo. In the water,two boys begin treading water and staying in the same place. Finally,the ball is thrown into the goal and the video pans out advertising the URL."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_VrNHEv6aR38.mp4",
"activitynet/videos/v_GgAXP4FTFnA.mp4",
"activitynet/videos/v_Wrbf7c58IuU.mp4",
"activitynet/videos/v_FLImHIKzzm4.mkv",
"activitynet/videos/v_8Ny9NjNpQQA.mp4"
] | 5
|
<answer>5</answer>
|
[
5
] | true
| 1
| 11,976.045533
| 185
|
[
{
"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 sitting outside on the bleachers watching a game of water polo. In the water,two boys begin treading water and staying in the same place. Finally,the ball is thrown into the goal and the video pans out advertising the URL.\"\n\nCandidates:\n[1] 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[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_GgAXP4FTFnA.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_Wrbf7c58IuU.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_FLImHIKzzm4.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_8Ny9NjNpQQA.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 child is playing on a playground. He crawls along a blue and red mat. He walks down a small green slide into a pit of balls."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_U_FxyViYYBA.mkv",
"activitynet/videos/v_bc-DycGxV9E.mp4",
"activitynet/videos/v_CCRPXH8ui-s.mp4",
"activitynet/videos/v_xpKAvKrrBDs.mp4",
"activitynet/videos/v_ibKFezOKsBQ.mp4"
] | 4
|
<answer>1</answer>
|
[
1
] | false
| 0
| 20,568.101872
| 184
|
[
{
"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 child is playing on a playground. He crawls along a blue and red mat. He walks down a small green slide into a pit of balls.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_U_FxyViYYBA.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_bc-DycGxV9E.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_CCRPXH8ui-s.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_xpKAvKrrBDs.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_ibKFezOKsBQ.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: "There are two men standing outside talking one is shirtless and he goes to grab some hair clippers. He stands in front of a mirror and starts to shave his bread off. He goes over one side of his face and then starts to shave his mustache a little. Then he returns so shaving his bread, he already looks very different."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_1VBg21aaiKM.mp4",
"activitynet/videos/v_qokr0bO828E.mp4",
"activitynet/videos/v_8_jbsmj5Z9w.mp4",
"activitynet/videos/v__3I4nm2zF5Y.mp4",
"activitynet/videos/v_s0swzu1jIpc.mkv"
] | 3
|
<answer>2</answer>
|
[
2
] | false
| 0
| 20,032.081079
| 186
|
[
{
"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: \"There are two men standing outside talking one is shirtless and he goes to grab some hair clippers. He stands in front of a mirror and starts to shave his bread off. He goes over one side of his face and then starts to shave his mustache a little. Then he returns so shaving his bread, he already looks very different.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_1VBg21aaiKM.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_qokr0bO828E.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_jbsmj5Z9w.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__3I4nm2zF5Y.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_s0swzu1jIpc.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 series of men run and attempt to pole vault but knock over the pole. Men pole vault and make it over the pole. We see the scores on a chain link fence. A man makes it over the pole and stands up and pumps his fists and we see the scores on the fence. A man makes it over the pole and a crowd stands and cheers. We see a red ending title screen."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_5rw6n16ILgY.mp4",
"activitynet/videos/v_U0qUFAPUg_Y.mp4",
"activitynet/videos/v_AZx_lm2XLHk.mp4",
"activitynet/videos/v_FrkXeG1YoKg.mp4",
"activitynet/videos/v_AJ15GW-sS5M.mp4"
] | 2
|
<answer>2</answer>
|
[
2
] | true
| 1
| 20,659.564369
| 187
|
[
{
"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 series of men run and attempt to pole vault but knock over the pole. Men pole vault and make it over the pole. We see the scores on a chain link fence. A man makes it over the pole and stands up and pumps his fists and we see the scores on the fence. A man makes it over the pole and a crowd stands and cheers. We see a red 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_5rw6n16ILgY.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_U0qUFAPUg_Y.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_AZx_lm2XLHk.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_FrkXeG1YoKg.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_AJ15GW-sS5M.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 chef adjust the cake pan into place and feels the top of a cake to check firmness. The chef uses a spatula to spread icing over the cake. The chef uses a ruler to guide a knife scoring the icing. The chef cuts the cake along the score lines and the finished cake is seen in pieces."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_QSV7f5XHohE.mkv",
"activitynet/videos/v_3Y_4Azzta6Q.mp4",
"activitynet/videos/v_u5ri43qbi1A.mp4",
"activitynet/videos/v_ss6XN-JP_x8.mp4",
"activitynet/videos/v_t7WI6H6UVG4.mkv"
] | 5
|
<answer>5</answer>
|
[
5
] | true
| 1
| 15,608.346523
| 188
|
[
{
"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 chef adjust the cake pan into place and feels the top of a cake to check firmness. The chef uses a spatula to spread icing over the cake. The chef uses a ruler to guide a knife scoring the icing. The chef cuts the cake along the score lines and the finished cake is seen in pieces.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_QSV7f5XHohE.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_3Y_4Azzta6Q.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_u5ri43qbi1A.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_ss6XN-JP_x8.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_t7WI6H6UVG4.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 girls are on a sandy beach. They are engaged in a game of volleyball. They lob the ball back and forth over the net."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_No5ZwqHdEQU.mp4",
"activitynet/videos/v_SnZnAVuMn4M.mp4",
"activitynet/videos/v_otq24Pdm3sc.mp4",
"activitynet/videos/v_laeOL4ipHck.mp4",
"activitynet/videos/v_UgXPt2LydrY.mp4"
] | 4
|
<answer>1</answer>
|
[
1
] | false
| 0
| 19,553.087673
| 189
|
[
{
"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 girls are on a sandy beach. They are engaged in a game of volleyball. They lob the ball back and forth over the net.\"\n\nCandidates:\n[1] 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[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_SnZnAVuMn4M.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_otq24Pdm3sc.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_laeOL4ipHck.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_UgXPt2LydrY.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's legs and feet are shown wearing untied ballerina shoes. The girl bends down and ties the ballerina shoes. The girl presents her tied ballerina shoes to the camera."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_Ld2a5ogu9k8.mp4",
"activitynet/videos/v_BCdt22s9hlU.mp4",
"activitynet/videos/v_ayLeSjJz53I.mkv",
"activitynet/videos/v_gYqXtgtyFnY.mp4",
"activitynet/videos/v_y3Zq6RZZNtc.mp4"
] | 4
|
<answer>4</answer>
|
[
4
] | true
| 1
| 17,771.11504
| 191
|
[
{
"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's legs and feet are shown wearing untied ballerina shoes. The girl bends down and ties the ballerina shoes. The girl presents her tied ballerina shoes to 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_Ld2a5ogu9k8.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_BCdt22s9hlU.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_ayLeSjJz53I.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_gYqXtgtyFnY.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_y3Zq6RZZNtc.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 putting chalk on his hands. He walks up to a large weight and picks it up over his head. He drops the weight and walks away to hug people watching."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_i1PpX1IOcIs.mp4",
"activitynet/videos/v_PY6WgOIZlhw.mp4",
"activitynet/videos/v_LCe0toF3058.mp4",
"activitynet/videos/v_AyicWbHhUWc.mp4",
"activitynet/videos/v_BzxK6r4UG5k.mp4"
] | 5
|
<answer>4</answer>
|
[
4
] | false
| 0
| 17,772.529743
| 190
|
[
{
"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 putting chalk on his hands. He walks up to a large weight and picks it up over his head. He drops the weight and walks away to hug people watching.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_i1PpX1IOcIs.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_PY6WgOIZlhw.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_LCe0toF3058.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_AyicWbHhUWc.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_BzxK6r4UG5k.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 lady petting a horse and talking. She shows a brush and starts brushing the horses mane. She takes a comb and brush and brushes the horses tail."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_ND9mMyNjm5M.mp4",
"activitynet/videos/v_FmLxc-aNh88.mkv",
"activitynet/videos/v_vBOejU7dBzY.mp4",
"activitynet/videos/v_0w7cO4tscBc.mp4",
"activitynet/videos/v_-MldnTjJ-zE.mp4"
] | 4
|
<answer>1</answer>
|
[
1
] | false
| 0
| 10,685.666676
| 193
|
[
{
"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 lady petting a horse and talking. She shows a brush and starts brushing the horses mane. She takes a comb and brush and brushes the horses tail.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_ND9mMyNjm5M.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_FmLxc-aNh88.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_vBOejU7dBzY.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_0w7cO4tscBc.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_-MldnTjJ-zE.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 suprises a woman in the bathroom while she is brushing her teeth in a bathroom. A toothbrush appears on screen while brushing a set of false teeth, the words "Modified Bass" are overlaid. The man talks while the woman brushes her teeth. The toothbrush returns to brush the false teeth. The man continues to talk while the woman brushes her teeth. A video subscribe screen appears for Coma Niddy. The white logo "PBS|digital studios" appears on a black screen."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_rWDMssiL7hE.mp4",
"activitynet/videos/v_6PnPu_cLCvE.mp4",
"activitynet/videos/v_pIk_bbjCNDo.mp4",
"activitynet/videos/v_gxILsv1RTEI.mp4",
"activitynet/videos/v_le1aEgEms9Y.mp4"
] | 1
|
<answer>2</answer>
|
[
2
] | false
| 0
| 15,714.026778
| 192
|
[
{
"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 suprises a woman in the bathroom while she is brushing her teeth in a bathroom. A toothbrush appears on screen while brushing a set of false teeth, the words \"Modified Bass\" are overlaid. The man talks while the woman brushes her teeth. The toothbrush returns to brush the false teeth. The man continues to talk while the woman brushes her teeth. A video subscribe screen appears for Coma Niddy. The white logo \"PBS|digital studios\" appears on a black screen.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_rWDMssiL7hE.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_6PnPu_cLCvE.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_pIk_bbjCNDo.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_gxILsv1RTEI.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_le1aEgEms9Y.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 gives a tutorial on how to change a bike tube. The wheel is taken off, tire irons are put in, and the tube is removed. A new tube is inserted and is pumped up."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_mSonugqhYuE.mkv",
"activitynet/videos/v_xIhTY02lRSE.mp4",
"activitynet/videos/v_XjV0D7nJx0Q.mp4",
"activitynet/videos/v_dRqbDamDLT0.mp4",
"activitynet/videos/v_PJgB6h-fImY.mp4"
] | 4
|
<answer>2</answer>
|
[
2
] | false
| 0
| 15,396.92869
| 194
|
[
{
"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 gives a tutorial on how to change a bike tube. The wheel is taken off, tire irons are put in, and the tube is removed. A new tube is inserted and is pumped up.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_mSonugqhYuE.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_xIhTY02lRSE.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_XjV0D7nJx0Q.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_dRqbDamDLT0.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_PJgB6h-fImY.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 takes a ball attached to a rope and swings it around his body. He throws the ball in the air. It lands on a field and people measure the distance. The man walks off in the other direction. It shows a replay of the man throwing the ball. The man walks over to more people. The track is shown in full view."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_e2QVdX-JdIg.mp4",
"activitynet/videos/v_zBm3FR-CCI0.mkv",
"activitynet/videos/v_9Zy5ylJYiA4.mp4",
"activitynet/videos/v_6cPXFUqRB1s.mp4",
"activitynet/videos/v_LAZHNzFbDNY.mp4"
] | 4
|
<answer>2</answer>
|
[
2
] | false
| 0
| 10,925.935794
| 196
|
[
{
"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 takes a ball attached to a rope and swings it around his body. He throws the ball in the air. It lands on a field and people measure the distance. The man walks off in the other direction. It shows a replay of the man throwing the ball. The man walks over to more people. The track is shown in full view.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_e2QVdX-JdIg.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_zBm3FR-CCI0.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_9Zy5ylJYiA4.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_6cPXFUqRB1s.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_LAZHNzFbDNY.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 throws a ball down an ice rink. A man in a red shirt is talking to the camera. People are scrubbing the ice in front of a ball."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_Oa26_SgrY8w.mp4",
"activitynet/videos/v_MP31A6fHsh4.mp4",
"activitynet/videos/v_le0tJsyuPks.mp4",
"activitynet/videos/v_9XmzbuByY_E.mp4",
"activitynet/videos/v_jDeBuorU4hY.mp4"
] | 2
|
<answer>2</answer>
|
[
2
] | true
| 1
| 11,646.709349
| 197
|
[
{
"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 throws a ball down an ice rink. A man in a red shirt is talking to the camera. People are scrubbing the ice in front of 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_Oa26_SgrY8w.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_MP31A6fHsh4.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_le0tJsyuPks.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_9XmzbuByY_E.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 large waterfall is shown on a rapid river. A person kayaks down the side of the fall, later being joined by others. They go speedily through the water, then a woman is shown speaking to the camera. We then watch several people as they kayak through the turbulent waters."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_vLL-voBPWM4.mp4",
"activitynet/videos/v_xs5imfBbWmw.mp4",
"activitynet/videos/v_k_xDTGiDp9A.mp4",
"activitynet/videos/v_gbuRv8phs1Y.mp4",
"activitynet/videos/v_f6wAW0Jv2Eo.mkv"
] | 4
|
<answer>2</answer>
|
[
2
] | false
| 0
| 29,158.903966
| 195
|
[
{
"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 waterfall is shown on a rapid river. A person kayaks down the side of the fall, later being joined by others. They go speedily through the water, then a woman is shown speaking to the camera. We then watch several people as they kayak through the turbulent waters.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_vLL-voBPWM4.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_xs5imfBbWmw.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_k_xDTGiDp9A.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_gbuRv8phs1Y.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_f6wAW0Jv2Eo.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 athlete is seen beginning to lift a heavy bar over his shoulders while moving in slow motion. The man continues to lift the bar over his head while jumping an finally throwing it down in the end."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_SKtUq_1cOSs.mp4",
"activitynet/videos/v_kNUpypAppjk.mp4",
"activitynet/videos/v_4YJ_L7jqgoA.mp4",
"activitynet/videos/v_PJysE5c1WDs.mp4",
"activitynet/videos/v_F67zl57FSXE.mkv"
] | 3
|
<answer>1</answer>
|
[
1
] | false
| 0
| 21,636.666426
| 198
|
[
{
"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 athlete is seen beginning to lift a heavy bar over his shoulders while moving in slow motion. The man continues to lift the bar over his head while jumping an finally throwing it down in the end.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_SKtUq_1cOSs.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_kNUpypAppjk.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_4YJ_L7jqgoA.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_PJysE5c1WDs.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_F67zl57FSXE.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 purple shirt is standing in a yard. There are three dogs in the yard. The dogs are playing with a frisbee."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_-Jp86pFKlsw.mp4",
"activitynet/videos/v_6nMQRUhOcwM.mp4",
"activitynet/videos/v_hSlydQ9rJuk.mp4",
"activitynet/videos/v__b_9BQvJ_v4.mp4",
"activitynet/videos/v_-qcPtBHelmc.mp4"
] | 1
|
<answer>3</answer>
|
[
3
] | false
| 0
| 22,395.089534
| 199
|
[
{
"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 purple shirt is standing in a yard. There are three dogs in the yard. The dogs are playing with a frisbee.\"\n\nCandidates:\n[1] 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
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_6nMQRUhOcwM.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_hSlydQ9rJuk.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_9BQvJ_v4.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_-qcPtBHelmc.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.