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 man and woman are spinning as they dance in a living room. They take turns, dancing in circles and spinning. They come to a stop and laugh."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_tm_CL7A0W4M.mp4",
"activitynet/videos/v_5Q_FrGFVGNY.mkv",
"activitynet/videos/v_qAZStAHJ3CQ.mp4",
"activitynet/videos/v_tjEMbP2SODQ.mp4",
"activitynet/videos/v_xf_exEkpJe8.mkv"
] | 2
|
<answer>2</answer>
|
[
2
] | true
| 1
| 18,939.609796
| 301
|
[
{
"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 woman are spinning as they dance in a living room. They take turns, dancing in circles and spinning. They come to a stop and laugh.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_tm_CL7A0W4M.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_5Q_FrGFVGNY.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_qAZStAHJ3CQ.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_tjEMbP2SODQ.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_xf_exEkpJe8.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A woman sits in her blue boat. The woman tugs on a rope. The woman begins accelerating her boat through the waters. Again, the woman tugs on the rope. The woman accelerates her little boat on the waters. The woman demonstrates tugging on the rope again. The woman accelerates her boat. For a final time, the woman tugs on the rope in her boat. The woman accelerates her boat one last time."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_7OYvyg32iqw.mp4",
"activitynet/videos/v_Mz-yz0fQ_Hk.mp4",
"activitynet/videos/v_kdfJW8YV378.mp4",
"activitynet/videos/v_W3TQnn0q9kc.mp4",
"activitynet/videos/v_GtCXZRGSaqk.mp4"
] | 2
|
<answer>2</answer>
|
[
2
] | true
| 1
| 25,481.202545
| 299
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames.\n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A woman sits in her blue boat. The woman tugs on a rope. The woman begins accelerating her boat through the waters. Again, the woman tugs on the rope. The woman accelerates her little boat on the waters. The woman demonstrates tugging on the rope again. The woman accelerates her boat. For a final time, the woman tugs on the rope in her boat. The woman accelerates her boat one last time.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_7OYvyg32iqw.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_Mz-yz0fQ_Hk.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_kdfJW8YV378.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_W3TQnn0q9kc.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_GtCXZRGSaqk.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 standing in a field throwing a frisbee around with a dog. The man does jumps and flips with the dog while still throwing the frisbee. The man continues to throw the object around with the dog jumping on him in the end."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_Cx3QGeQu7xM.mp4",
"activitynet/videos/v_fh68-PXZ9Oo.mp4",
"activitynet/videos/v_f3YyN44Dx8M.mp4",
"activitynet/videos/v_NK5FWZ2BOQs.mp4",
"activitynet/videos/v__b_9BQvJ_v4.mp4"
] | 2
|
<answer>1</answer>
|
[
1
] | false
| 0
| 16,230.043715
| 302
|
[
{
"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 standing in a field throwing a frisbee around with a dog. The man does jumps and flips with the dog while still throwing the frisbee. The man continues to throw the object around with the dog jumping on him 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_Cx3QGeQu7xM.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_fh68-PXZ9Oo.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_f3YyN44Dx8M.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_NK5FWZ2BOQs.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__b_9BQvJ_v4.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: "ingredients are shown on top of table with bowls. cake mix in bowl with water and oil and eggs and is mixed till moistened and is divided in bowls and poured into pan and in the oven. when is ready is combined with frost and candies on top."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_WFL8DhccHr0.mp4",
"activitynet/videos/v_Uqs8NaPzHKU.mp4",
"activitynet/videos/v_yyCsQ7QzAJ8.mp4",
"activitynet/videos/v_gLfIPN_WM48.mp4",
"activitynet/videos/v_nKa1e_CpvoY.mp4"
] | 1
|
<answer>2</answer>
|
[
2
] | false
| 0
| 23,675.129443
| 303
|
[
{
"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: \"ingredients are shown on top of table with bowls. cake mix in bowl with water and oil and eggs and is mixed till moistened and is divided in bowls and poured into pan and in the oven. when is ready is combined with frost and candies on top.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_WFL8DhccHr0.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_Uqs8NaPzHKU.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_yyCsQ7QzAJ8.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_gLfIPN_WM48.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_nKa1e_CpvoY.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 snow gear comes walking up and puts his skis on to go skiing down the snow. He starts, holding a ticker in one hand and bringing his body down so that he picks up more speed. Once he's going fast enough he stands up and then squats again, he seems to have a technique that works for him. He continues down the snowy slope skiing on his own."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_H-5nHSHwFOk.mp4",
"activitynet/videos/v_m_gr7WdjJmc.mp4",
"activitynet/videos/v_0y_5NIIvUzI.mp4",
"activitynet/videos/v_7H5oYHs7EJw.mp4",
"activitynet/videos/v_iMATWwGyAUM.mp4"
] | 4
|
<answer>2</answer>
|
[
2
] | false
| 0
| 31,812.530594
| 304
|
[
{
"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 snow gear comes walking up and puts his skis on to go skiing down the snow. He starts, holding a ticker in one hand and bringing his body down so that he picks up more speed. Once he's going fast enough he stands up and then squats again, he seems to have a technique that works for him. He continues down the snowy slope skiing on his own.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_H-5nHSHwFOk.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_m_gr7WdjJmc.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_0y_5NIIvUzI.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_7H5oYHs7EJw.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_iMATWwGyAUM.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 dance team dressed in black with pink letters shows a routine put to reggae music. The women jump bounce in place before the lyrics of the song starts. When the lyrics start they go into an intricate dance routine that fits with the song. After the song is over the group of winded ladies disbands from the formation."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_j4-w606GnYw.mp4",
"activitynet/videos/v_KhAtzEJxz9M.mp4",
"activitynet/videos/v_jRXF5_vNUWE.mp4",
"activitynet/videos/v_K3Z3z8t-RIQ.mp4",
"activitynet/videos/v_m-B1tlnywNY.mp4"
] | 2
|
<answer>2</answer>
|
[
2
] | true
| 1
| 27,929.112598
| 305
|
[
{
"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 dance team dressed in black with pink letters shows a routine put to reggae music. The women jump bounce in place before the lyrics of the song starts. When the lyrics start they go into an intricate dance routine that fits with the song. After the song is over the group of winded ladies disbands from the formation.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_j4-w606GnYw.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_KhAtzEJxz9M.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_jRXF5_vNUWE.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_K3Z3z8t-RIQ.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_m-B1tlnywNY.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A man demonstrates, in a living room setting, how to play the harmonica. A man holds a harmonica up to the camera in a close up shot and then begins to play using both hands to cover the harmonica while he plays. The man continues to play the harmonica, moving back and forth and side to side and emoting with his eyebrows while playing. The man stops playing and begins to talk to the camera using hand gestures to point to the harmonica, at times. The man then begins playing again using his hands to indicate the type of note he is hitting. The man then begins playing again before stopping and talking to the camera one last time."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_jVC3DZdphYM.mp4",
"activitynet/videos/v_B_PhHrBEeNI.mp4",
"activitynet/videos/v_Ue90f5r-2Qw.mp4",
"activitynet/videos/v_lDJpGI4BZ8k.mkv",
"activitynet/videos/v_fKDl_CnA8nY.mp4"
] | 1
|
<answer>1</answer>
|
[
1
] | true
| 1
| 12,676.903321
| 306
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames.\n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A man demonstrates, in a living room setting, how to play the harmonica. A man holds a harmonica up to the camera in a close up shot and then begins to play using both hands to cover the harmonica while he plays. The man continues to play the harmonica, moving back and forth and side to side and emoting with his eyebrows while playing. The man stops playing and begins to talk to the camera using hand gestures to point to the harmonica, at times. The man then begins playing again using his hands to indicate the type of note he is hitting. The man then begins playing again before stopping and talking to the camera one last time.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_jVC3DZdphYM.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_B_PhHrBEeNI.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_Ue90f5r-2Qw.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_lDJpGI4BZ8k.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_fKDl_CnA8nY.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 scuba diving under the blue water. Colorful corals and school of fish are being shown. The scuba divers continue to swim as the view of colorful corals are being shown."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_QuEHZ2Y3H40.mp4",
"activitynet/videos/v_Dydb923dXss.mp4",
"activitynet/videos/v_P1gGM89_T2g.mkv",
"activitynet/videos/v_W30cufYc_ZI.mp4",
"activitynet/videos/v_61wzTjdnXe4.mp4"
] | 2
|
<answer>2</answer>
|
[
2
] | true
| 1
| 28,868.307059
| 308
|
[
{
"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 scuba diving under the blue water. Colorful corals and school of fish are being shown. The scuba divers continue to swim as the view of colorful corals are being shown.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_QuEHZ2Y3H40.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_Dydb923dXss.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_P1gGM89_T2g.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_W30cufYc_ZI.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_61wzTjdnXe4.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 camera shows a rough lake and landscapes as well as people off in the distance and a man giving the camera a thumbs up. The camera zooms in on people riding along in tubes."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_3jZq0UaDIks.mp4",
"activitynet/videos/v_yJSQmNSFlNI.mkv",
"activitynet/videos/v_SMNXIkCGh_0.mp4",
"activitynet/videos/v_BKdKbFPerGo.mp4",
"activitynet/videos/v_bG55LSFBA9M.mp4"
] | 3
|
<answer>2</answer>
|
[
2
] | false
| 0
| 34,316.757583
| 307
|
[
{
"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 camera shows a rough lake and landscapes as well as people off in the distance and a man giving the camera a thumbs up. The camera zooms in on people riding along in tubes.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_3jZq0UaDIks.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_yJSQmNSFlNI.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_SMNXIkCGh_0.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_BKdKbFPerGo.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_bG55LSFBA9M.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 walking around on the snow. People are going down the hill of snow in tubes. People are coming up the hill attached to a rope."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_5kmGgH4xFW0.mp4",
"activitynet/videos/v_g4uvBcIE1Os.mp4",
"activitynet/videos/v_vdq_xoRyxCU.mp4",
"activitynet/videos/v_OMRu1rPRBHo.mp4",
"activitynet/videos/v_HzAlvJ1fNWU.mp4"
] | 4
|
<answer>2</answer>
|
[
2
] | false
| 0
| 34,737.903757
| 309
|
[
{
"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 walking around on the snow. People are going down the hill of snow in tubes. People are coming up the hill attached to a rope.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_5kmGgH4xFW0.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_g4uvBcIE1Os.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_vdq_xoRyxCU.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_OMRu1rPRBHo.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_HzAlvJ1fNWU.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 playing a guitar while another man cuts the hair of a boy behind him. The man continues playing as another man brushes a comb through the boy's hair."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_9WhPG89P-tg.mp4",
"activitynet/videos/v_Y5zJT3BjIxM.mp4",
"activitynet/videos/v_CIgdBoHjGXU.mp4",
"activitynet/videos/v_0ZHZ1ZqmT7s.mp4",
"activitynet/videos/v_yl37hI-Bgkk.mp4"
] | 4
|
<answer>4</answer>
|
[
4
] | true
| 1
| 17,876.964732
| 310
|
[
{
"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 playing a guitar while another man cuts the hair of a boy behind him. The man continues playing as another man brushes a comb through the boy's hair.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_9WhPG89P-tg.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_Y5zJT3BjIxM.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_CIgdBoHjGXU.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_0ZHZ1ZqmT7s.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_yl37hI-Bgkk.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A person is seen riding down a hill on a skateboard with another person following behind. More clips are shown of people riding down a hill on skateboards as well as luging with others. More clips are shown of people riding down as well as wiping out afterwards and the camera capturing their movements."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_ywSeEtroEXo.mp4",
"activitynet/videos/v_nzjivjnk2Ac.mp4",
"activitynet/videos/v_oO1g33vi4hg.mp4",
"activitynet/videos/v_X5UoLcloHIM.mp4",
"activitynet/videos/v_iZk3PH8ghlI.mp4"
] | 1
|
<answer>5</answer>
|
[
5
] | false
| 0
| 20,322.506227
| 311
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames.\n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A person is seen riding down a hill on a skateboard with another person following behind. More clips are shown of people riding down a hill on skateboards as well as luging with others. More clips are shown of people riding down as well as wiping out afterwards and the camera capturing their movements.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_ywSeEtroEXo.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_nzjivjnk2Ac.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_oO1g33vi4hg.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_X5UoLcloHIM.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_iZk3PH8ghlI.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 dawn scene is shown as two people walk down the road. There is a building made of bricks, and a man with an array of predecorated pumpkins. He talks to the camera, then is shown handing and stacking the pumpkins for the public. A woman talks about him as he carves the pumpkins and people watch. At the end, a pumpkin with the words we miss you edgar are shown, indicating the man had passed away."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_gyQ3NBwXhDU.mp4",
"activitynet/videos/v_PSh-caJvSHU.mp4",
"activitynet/videos/v_TjR436qaQw4.mp4",
"activitynet/videos/v_biyf6Q-xF0M.mp4",
"activitynet/videos/v_C1IuvUSmcvA.mp4"
] | 4
|
<answer>4</answer>
|
[
4
] | true
| 1
| 24,753.23343
| 312
|
[
{
"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 dawn scene is shown as two people walk down the road. There is a building made of bricks, and a man with an array of predecorated pumpkins. He talks to the camera, then is shown handing and stacking the pumpkins for the public. A woman talks about him as he carves the pumpkins and people watch. At the end, a pumpkin with the words we miss you edgar are shown, indicating the man had passed away.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_gyQ3NBwXhDU.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_PSh-caJvSHU.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_TjR436qaQw4.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_biyf6Q-xF0M.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_C1IuvUSmcvA.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A person is seen with a dog in a large open field. The person throws a frisbee off in the distance and the dog chases after it. In the end the camera falls over ad he walks over to grab it."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_KRES3eBM2l4.mp4",
"activitynet/videos/v_hSlydQ9rJuk.mp4",
"activitynet/videos/v_1XtjXqqPvyQ.mp4",
"activitynet/videos/v_Fi2Al65EH0g.mp4",
"activitynet/videos/v__b_9BQvJ_v4.mp4"
] | 1
|
<answer>2</answer>
|
[
2
] | false
| 0
| 17,151.557779
| 314
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames.\n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A person is seen with a dog in a large open field. The person throws a frisbee off in the distance and the dog chases after it. In the end the camera falls over ad he walks over to grab it.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_KRES3eBM2l4.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_hSlydQ9rJuk.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_1XtjXqqPvyQ.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_Fi2Al65EH0g.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__b_9BQvJ_v4.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 clips of girls performing dance moves are shown on a large dance floor with a small group of judges sitting behind a table watching them. The girls continue spinning on the floor and leads into them performing moves at a barre and a man instructing them on the side."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_dTwH5Fzu4eE.mp4",
"activitynet/videos/v_HDVk1O78gwc.mp4",
"activitynet/videos/v_tznMNEWglxY.mp4",
"activitynet/videos/v_2Iakg-Z-iXM.mkv",
"activitynet/videos/v_Fok7z0mLNbU.mkv"
] | 5
|
<answer>5</answer>
|
[
5
] | true
| 1
| 15,010.717844
| 315
|
[
{
"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 clips of girls performing dance moves are shown on a large dance floor with a small group of judges sitting behind a table watching them. The girls continue spinning on the floor and leads into them performing moves at a barre and a man instructing them on the side.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_dTwH5Fzu4eE.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_HDVk1O78gwc.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_tznMNEWglxY.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_2Iakg-Z-iXM.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_Fok7z0mLNbU.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 underside of a canoe and paddle is shown. A man is paddling the canoe down a river. He talks as he canoes, pulling it through shallow areas of water and rocks before getting back in. The man drinks and eats in the canoe, then continues down river."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_h67ctuwV-Nw.mp4",
"activitynet/videos/v_okvQJRTfGHk.mp4",
"activitynet/videos/v_Liha_xwiwtc.mp4",
"activitynet/videos/v_RpyIg_j4I3E.mkv",
"activitynet/videos/v_Ff8QLpH5T1c.mkv"
] | 5
|
<answer>5</answer>
|
[
5
] | true
| 1
| 34,141.00142
| 313
|
[
{
"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 underside of a canoe and paddle is shown. A man is paddling the canoe down a river. He talks as he canoes, pulling it through shallow areas of water and rocks before getting back in. The man drinks and eats in the canoe, then continues down river.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_h67ctuwV-Nw.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_okvQJRTfGHk.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_Liha_xwiwtc.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_RpyIg_j4I3E.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_Ff8QLpH5T1c.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A young woman in a black shirt and jean pants is sitting on the back of a grey horse with a long grey mane in the middle of a large grassy field with hills behind them. The horse begins to trot and then slows down a little bit and starts smiling as she steers the horse in a circle. As they return to their original spot,you can see the water washing up against the land and they continue back up the hill."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_2GACaR0GdD8.mp4",
"activitynet/videos/v_2aHetC-N-P4.mp4",
"activitynet/videos/v_fO8b3U8fuGo.mp4",
"activitynet/videos/v_Jz7bt59z6Qg.mp4",
"activitynet/videos/v_qXNYHbnGvto.mp4"
] | 3
|
<answer>4</answer>
|
[
4
] | false
| 0
| 13,969.837886
| 316
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames.\n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A young woman in a black shirt and jean pants is sitting on the back of a grey horse with a long grey mane in the middle of a large grassy field with hills behind them. The horse begins to trot and then slows down a little bit and starts smiling as she steers the horse in a circle. As they return to their original spot,you can see the water washing up against the land and they continue back up the hill.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_2GACaR0GdD8.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_2aHetC-N-P4.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_fO8b3U8fuGo.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_Jz7bt59z6Qg.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_qXNYHbnGvto.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 guy leaves a crowd, chases after a black cattle, lassoes the cattle. The guy gets of his horse and ties the cattle's legs. The guy gets on his horse. A man on a dark horse rides past the cattle. The cattle tries to raise."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_l2MB-KxbVEs.mp4",
"activitynet/videos/v_LA5UXJ_hVU4.mp4",
"activitynet/videos/v_vBKIXqRd-eA.mkv",
"activitynet/videos/v_2QydTDAYJsA.mp4",
"activitynet/videos/v_6f1HnAlpphA.mp4"
] | 5
|
<answer>1</answer>
|
[
1
] | false
| 0
| 12,606.428741
| 317
|
[
{
"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 guy leaves a crowd, chases after a black cattle, lassoes the cattle. The guy gets of his horse and ties the cattle's legs. The guy gets on his horse. A man on a dark horse rides past the cattle. The cattle tries to raise.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_l2MB-KxbVEs.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_LA5UXJ_hVU4.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_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_2QydTDAYJsA.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_6f1HnAlpphA.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 guy carries a child on an inflated tire, and they ride the rocky body of water. The child falls into the water but gets back on the guy's lap. The inflated tire turns over and both enter the body of water."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_3OcAjx8e4LU.mp4",
"activitynet/videos/v_pbz8c7TAlDs.mkv",
"activitynet/videos/v_Fr9F2xRLd0A.mp4",
"activitynet/videos/v_-u2zAMnrCC4.mp4",
"activitynet/videos/v_u-X4YO91V78.mkv"
] | 2
|
<answer>2</answer>
|
[
2
] | true
| 1
| 16,978.171531
| 318
|
[
{
"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 guy carries a child on an inflated tire, and they ride the rocky body of water. The child falls into the water but gets back on the guy's lap. The inflated tire turns over and both enter the body of water.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_3OcAjx8e4LU.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_pbz8c7TAlDs.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_Fr9F2xRLd0A.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_-u2zAMnrCC4.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_u-X4YO91V78.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 and boy are washing dishes in a kitchen. The camera moves to in front of the kids and their mom enters the room. The mom adds something to the fridge and puts something in a cabinet. The mother leaves the room. the mother enters and adds some dishes from the rack back into the sink to be rinsed again and shows the boy what was wrong with the pot."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_P6Z-7k-erfc.mp4",
"activitynet/videos/v_kuMevlNUDCs.mp4",
"activitynet/videos/v_DinaQYSgbtg.mp4",
"activitynet/videos/v_iY7bZQnHXlk.mp4",
"activitynet/videos/v_u8ykXBc2Efs.mkv"
] | 5
|
<answer>5</answer>
|
[
5
] | true
| 1
| 13,162.953944
| 319
|
[
{
"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 and boy are washing dishes in a kitchen. The camera moves to in front of the kids and their mom enters the room. The mom adds something to the fridge and puts something in a cabinet. The mother leaves the room. the mother enters and adds some dishes from the rack back into the sink to be rinsed again and shows the boy what was wrong with the pot.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_P6Z-7k-erfc.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_kuMevlNUDCs.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_DinaQYSgbtg.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_iY7bZQnHXlk.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_u8ykXBc2Efs.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A lady is washing dishes in a kitchen. The camera moves to her right side. A man plays with her hair. The man touches her arm. The camera blurs on the girls face."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_uTQyPHg8r0M.mp4",
"activitynet/videos/v_CSDApI2nHPU.mp4",
"activitynet/videos/v_optJ47P_5Ys.mp4",
"activitynet/videos/v_xaicDAewb6o.mp4",
"activitynet/videos/v_64UBH371Jj8.mp4"
] | 5
|
<answer>2</answer>
|
[
2
] | false
| 0
| 8,808.267611
| 321
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames.\n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A lady is washing dishes in a kitchen. The camera moves to her right side. A man plays with her hair. The man touches her arm. The camera blurs on the girls face.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_uTQyPHg8r0M.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_CSDApI2nHPU.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_optJ47P_5Ys.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_xaicDAewb6o.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: "A woman puts a cookie in her mouth, then ingredients to make cookies are display on a table. The woman crush dates in a blender, then adds wet ingredients and oil and mix. Then, the woman takes out the dough and makes dough balls and puts chocolate chips inside."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_MVzypK0eMKc.mp4",
"activitynet/videos/v_cyhWzLsM29E.mp4",
"activitynet/videos/v_Hhc10CrukfA.mp4",
"activitynet/videos/v_HOTCR1uIaBM.mp4",
"activitynet/videos/v_Pv6oIFroaCQ.mp4"
] | 2
|
<answer>2</answer>
|
[
2
] | true
| 1
| 20,410.738291
| 320
|
[
{
"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 puts a cookie in her mouth, then ingredients to make cookies are display on a table. The woman crush dates in a blender, then adds wet ingredients and oil and mix. Then, the woman takes out the dough and makes dough balls and puts chocolate chips inside.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_MVzypK0eMKc.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_cyhWzLsM29E.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_Hhc10CrukfA.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_HOTCR1uIaBM.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_Pv6oIFroaCQ.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 practice basketball dribbling behind chairs. Then, the man advance with the ball and shoot it in the basket."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_K0MzjnMzbj4.mp4",
"activitynet/videos/v_S6VgTNGiIkg.mp4",
"activitynet/videos/v_sRol1BJ9EUk.mp4",
"activitynet/videos/v_1JKgr3KfoHo.mkv",
"activitynet/videos/v_ykov_joUUTk.mp4"
] | 3
|
<answer>3</answer>
|
[
3
] | true
| 1
| 18,494.360352
| 322
|
[
{
"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 practice basketball dribbling behind chairs. Then, the man advance with the ball and shoot it in the basket.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_K0MzjnMzbj4.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_S6VgTNGiIkg.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_sRol1BJ9EUk.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_1JKgr3KfoHo.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_ykov_joUUTk.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 person scraping ice off their windshield with the camera inside the car. A person makes a hand sign to the woman scraping the ice. The lady switches to the other side of the car. We see the man's face as he sits in the car."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_15npAlupNU4.mp4",
"activitynet/videos/v_OMa1i3ITBbo.mkv",
"activitynet/videos/v_R9qRR8CcSJA.mp4",
"activitynet/videos/v_MR0vMF_5hp8.mp4",
"activitynet/videos/v_n-BJ753InB0.mp4"
] | 4
|
<answer>5</answer>
|
[
5
] | false
| 0
| 14,680.81804
| 323
|
[
{
"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 person scraping ice off their windshield with the camera inside the car. A person makes a hand sign to the woman scraping the ice. The lady switches to the other side of the car. We see the man's face as he sits in the car.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_15npAlupNU4.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_OMa1i3ITBbo.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_R9qRR8CcSJA.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_MR0vMF_5hp8.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_n-BJ753InB0.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 green outfit, wearing white gloves and a green hat is bending over with his hands on his thighs while a website is displayed on the bottom of the screen. The man stands up and other men appear talking to him or are near him. The men are positioned on the field and a man hits a ball and and all the men start running on the field. The man who just hit the ball begins to cheer and jump up and down multiple times then walks away."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_7X3wPRKuAsU.mp4",
"activitynet/videos/v_9jivQgF6J1g.mp4",
"activitynet/videos/v_OuEQLjwBIPI.mp4",
"activitynet/videos/v_uyr3E9ZReAw.mkv",
"activitynet/videos/v_v5patZyuYys.mp4"
] | 3
|
<answer>4</answer>
|
[
4
] | false
| 0
| 10,805.945298
| 324
|
[
{
"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 green outfit, wearing white gloves and a green hat is bending over with his hands on his thighs while a website is displayed on the bottom of the screen. The man stands up and other men appear talking to him or are near him. The men are positioned on the field and a man hits a ball and and all the men start running on the field. The man who just hit the ball begins to cheer and jump up and down multiple times then walks away.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_7X3wPRKuAsU.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_9jivQgF6J1g.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_OuEQLjwBIPI.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_uyr3E9ZReAw.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_v5patZyuYys.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 exercise on elliptical machines that are comfortable. A woman holds the large ergonomic handles and then the small handles while pedaling. The elliptical machine has a board to program the exercises. Then, a man moves an elliptical machine, and an old woman fix the top of the woman. A man takes a picture. The woman continues pedaling the bike."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_swbCsf51XVg.mp4",
"activitynet/videos/v_E15Q3Z9J-Zg.mp4",
"activitynet/videos/v_R3HC-IAZVZg.mp4",
"activitynet/videos/v_zpBZ7HMNO34.mp4",
"activitynet/videos/v_e6mpdQ3BFhA.mp4"
] | 5
|
<answer>2</answer>
|
[
2
] | false
| 0
| 14,877.488142
| 325
|
[
{
"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 exercise on elliptical machines that are comfortable. A woman holds the large ergonomic handles and then the small handles while pedaling. The elliptical machine has a board to program the exercises. Then, a man moves an elliptical machine, and an old woman fix the top of the woman. A man takes a picture. The woman continues pedaling the bike.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_swbCsf51XVg.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_E15Q3Z9J-Zg.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_R3HC-IAZVZg.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_zpBZ7HMNO34.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_e6mpdQ3BFhA.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 howcast logo appears and the title of the video, how to clean the bathroom sink and the dirty sink is shown. The items that are needed to wash the sink was listed and individually put in the sink. The cleaning starts with the step by step procedure and tips on how to clean the sink. An information showing how many bathrooms a certain mansion has and ended on howcast logo."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_t2DdSm_MGXo.mp4",
"activitynet/videos/v_1UgjxeAPq_A.mp4",
"activitynet/videos/v_-7eQ2bHNPUw.mkv",
"activitynet/videos/v_RnShLAifVno.mp4",
"activitynet/videos/v_R-RQx5pbMvo.mp4"
] | 1
|
<answer>3</answer>
|
[
3
] | false
| 0
| 14,573.091134
| 326
|
[
{
"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 howcast logo appears and the title of the video, how to clean the bathroom sink and the dirty sink is shown. The items that are needed to wash the sink was listed and individually put in the sink. The cleaning starts with the step by step procedure and tips on how to clean the sink. An information showing how many bathrooms a certain mansion has and ended on howcast logo.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_t2DdSm_MGXo.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_1UgjxeAPq_A.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_-7eQ2bHNPUw.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_RnShLAifVno.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_R-RQx5pbMvo.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 on bikes in a room. They all peddle in place. The camera pans back and forth between them. They keep peddling the whole time."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_qiRrR2Nj2SQ.mp4",
"activitynet/videos/v_nw32dno_RcM.mp4",
"activitynet/videos/v_lcEGoZAC7GI.mp4",
"activitynet/videos/v_R7iFa9OpoTY.mp4",
"activitynet/videos/v_Ttx3kt2fW1o.mp4"
] | 5
|
<answer>4</answer>
|
[
4
] | false
| 0
| 18,939.547245
| 327
|
[
{
"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 on bikes in a room. They all peddle in place. The camera pans back and forth between them. They keep peddling the whole time.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_qiRrR2Nj2SQ.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_nw32dno_RcM.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_lcEGoZAC7GI.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_R7iFa9OpoTY.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_Ttx3kt2fW1o.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 girl walking slowly with a shuffleboard pole and pushing a puck. the lady gets to the blue line and pushes the puck. the puck flips and sticks instead of sliding. the lady turns and walks away."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_USWExMIMcik.mp4",
"activitynet/videos/v_b4b6YkxsHk4.mp4",
"activitynet/videos/v_mfJj5gBQg-4.mkv",
"activitynet/videos/v_4Ex-sB0vtwk.mp4",
"activitynet/videos/v_9Xrw-WOipSI.mp4"
] | 2
|
<answer>2</answer>
|
[
2
] | true
| 1
| 9,922.965926
| 328
|
[
{
"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 girl walking slowly with a shuffleboard pole and pushing a puck. the lady gets to the blue line and pushes the puck. the puck flips and sticks instead of sliding. the lady turns and walks away.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_USWExMIMcik.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_b4b6YkxsHk4.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_mfJj5gBQg-4.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_4Ex-sB0vtwk.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_9Xrw-WOipSI.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 layup and passing drill is something that is done. The girls then do the passing and laying up."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_mUyMYnGXKgk.mp4",
"activitynet/videos/v_L21zcZlFfIY.mp4",
"activitynet/videos/v_FU0EPNGKsv8.mp4",
"activitynet/videos/v_LbXhdPZakpo.mp4",
"activitynet/videos/v_cICxG-28hK0.mp4"
] | 1
|
<answer>1</answer>
|
[
1
] | true
| 1
| 13,299.927269
| 331
|
[
{
"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 layup and passing drill is something that is done. The girls then do the passing and laying up.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_mUyMYnGXKgk.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_L21zcZlFfIY.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_FU0EPNGKsv8.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_LbXhdPZakpo.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_cICxG-28hK0.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 skiing down some snow at a very fast speed. Snow is building up all over his face, it looks extremely cold. He kind of almost falls down but continues moving. He is skiing fast through trees, passing by other skiers, he continues to go so fast."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_cY3QbnSeu9k.mp4",
"activitynet/videos/v_6VygM9-XgAk.mp4",
"activitynet/videos/v_jWPr92KwXeY.mp4",
"activitynet/videos/v_Vg043D46E7Q.mp4",
"activitynet/videos/v_iJqLgrShN-w.mp4"
] | 5
|
<answer>5</answer>
|
[
5
] | true
| 1
| 25,351.879411
| 330
|
[
{
"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 skiing down some snow at a very fast speed. Snow is building up all over his face, it looks extremely cold. He kind of almost falls down but continues moving. He is skiing fast through trees, passing by other skiers, he continues to go so fast.\"\n\nCandidates:\n[1] 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[2] 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
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_jWPr92KwXeY.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_Vg043D46E7Q.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_iJqLgrShN-w.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 wearing headphones while seated with a guitar. He is playing the guitar through a microphone located in front of him. He rocks the guitar back and forth and looks at the camera, smiling as he finishes."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_sARnRvNdl-Y.mp4",
"activitynet/videos/v_qNxLTF4Q6yk.mp4",
"activitynet/videos/v_a8dUtKcAunw.mp4",
"activitynet/videos/v_fJNauQt9Di0.mp4",
"activitynet/videos/v_y-rgla4aNUo.mp4"
] | 1
|
<answer>1</answer>
|
[
1
] | true
| 1
| 33,486.286556
| 329
|
[
{
"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 wearing headphones while seated with a guitar. He is playing the guitar through a microphone located in front of him. He rocks the guitar back and forth and looks at the camera, smiling as he finishes.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_sARnRvNdl-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_qNxLTF4Q6yk.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_a8dUtKcAunw.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_fJNauQt9Di0.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_y-rgla4aNUo.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A man is seen sitting in a tube drinking a beer while the camera pans around him and another sitting in a tube. A cooler is seen sitting in the middle as the man continues speaking and the camera pans around the water."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_BKdKbFPerGo.mp4",
"activitynet/videos/v_sG3JpMuXFnU.mp4",
"activitynet/videos/v_6DYQHmsezUw.mp4",
"activitynet/videos/v_jpclX7wgcZU.mp4",
"activitynet/videos/v_dzR4voNDZ7g.mp4"
] | 4
|
<answer>4</answer>
|
[
4
] | true
| 1
| 18,646.938724
| 332
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames.\n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A man is seen sitting in a tube drinking a beer while the camera pans around him and another sitting in a tube. A cooler is seen sitting in the middle as the man continues speaking and the camera pans around 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_BKdKbFPerGo.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_sG3JpMuXFnU.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_6DYQHmsezUw.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_jpclX7wgcZU.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_dzR4voNDZ7g.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 several shots of various men performing martial arts flips. The video leads into boys doing flips onto trampolines as well as large bans to catch their fall."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_CoP3xaSZt7A.mkv",
"activitynet/videos/v_kW4ajodPtWU.mp4",
"activitynet/videos/v_xbcP38aF5Ok.mp4",
"activitynet/videos/v_E2NKQZNMAO0.mp4",
"activitynet/videos/v__uOfIm1tFcI.mp4"
] | 5
|
<answer>5</answer>
|
[
5
] | true
| 1
| 12,883.438356
| 333
|
[
{
"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 several shots of various men performing martial arts flips. The video leads into boys doing flips onto trampolines as well as large bans to catch their fall.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_CoP3xaSZt7A.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_kW4ajodPtWU.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_xbcP38aF5Ok.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_E2NKQZNMAO0.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__uOfIm1tFcI.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 words skate like a pro are shown at the bottom. A man roller blades backwards. A close up of the skates are shown. A skater bends over as they skate."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_AuVVP8q6tFY.mp4",
"activitynet/videos/v_BDQHEemWnSk.mp4",
"activitynet/videos/v_XASTWKClhPU.mp4",
"activitynet/videos/v_2TEJnQzCPUM.mp4",
"activitynet/videos/v_9UpVdljXQ4E.mp4"
] | 5
|
<answer>5</answer>
|
[
5
] | true
| 1
| 19,260.238092
| 334
|
[
{
"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 words skate like a pro are shown at the bottom. A man roller blades backwards. A close up of the skates are shown. A skater bends over as they skate.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_AuVVP8q6tFY.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_BDQHEemWnSk.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_XASTWKClhPU.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_2TEJnQzCPUM.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_9UpVdljXQ4E.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 is seen bending forward with another child standing right behind her holding a net. The girl sticks various objects under some running water and the boy hands her a spoon to wash."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_XyQSmMYbP6o.mp4",
"activitynet/videos/v_xaicDAewb6o.mp4",
"activitynet/videos/v_8xvoAyY70I8.mp4",
"activitynet/videos/v_n4apOkL24BE.mp4",
"activitynet/videos/v_69DNcmkoapw.mp4"
] | 3
|
<answer>2</answer>
|
[
2
] | false
| 0
| 15,799.920475
| 335
|
[
{
"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 is seen bending forward with another child standing right behind her holding a net. The girl sticks various objects under some running water and the boy hands her a spoon to wash.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_XyQSmMYbP6o.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_xaicDAewb6o.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_8xvoAyY70I8.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_n4apOkL24BE.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_69DNcmkoapw.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 named Russ is standing in a shop called Out-N-Back Bicycle Specialist and is holding a tire as he talks. A few colorful screens display with words on them. The man is once again shown with the tire and he begins to show how he removes the rubber from the rim using a blue object. The man then removes the black rubber with just his hands and removes a metal screw piece, then replaces it once again by maneuvering it with his two hands. When the man is done he holds the tire upright and continues to talk. The outro begins and it is similiar to the colorful screens shown earlier, but ends with a white screen that has footprints printed on it and the words UTAH ADVENTURER dot COM."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_4sm-tTbfamM.mp4",
"activitynet/videos/v_-doxoUNGLJE.mp4",
"activitynet/videos/v_xIhTY02lRSE.mp4",
"activitynet/videos/v_XjV0D7nJx0Q.mp4",
"activitynet/videos/v_vdTisVMhW7I.mp4"
] | 4
|
<answer>4</answer>
|
[
4
] | true
| 1
| 20,665.796531
| 336
|
[
{
"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 named Russ is standing in a shop called Out-N-Back Bicycle Specialist and is holding a tire as he talks. A few colorful screens display with words on them. The man is once again shown with the tire and he begins to show how he removes the rubber from the rim using a blue object. The man then removes the black rubber with just his hands and removes a metal screw piece, then replaces it once again by maneuvering it with his two hands. When the man is done he holds the tire upright and continues to talk. The outro begins and it is similiar to the colorful screens shown earlier, but ends with a white screen that has footprints printed on it and the words UTAH ADVENTURER dot COM.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_4sm-tTbfamM.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_-doxoUNGLJE.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_xIhTY02lRSE.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_XjV0D7nJx0Q.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_vdTisVMhW7I.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 with a cup in hand next to a horse. The horse begins to drink from the man's cup. Some kids and some dogs stand buy and watch. The man begins to extend his arm offering the cup to the horse. The horse proceeds to like the cup dry."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_IXUh06YCtjw.mp4",
"activitynet/videos/v_rAO-_VxIJng.mp4",
"activitynet/videos/v_EiXW33yuAcw.mp4",
"activitynet/videos/v_6q3EIv2X8BQ.mp4",
"activitynet/videos/v_KFS_lGlO-Ew.mp4"
] | 3
|
<answer>3</answer>
|
[
3
] | true
| 1
| 10,917.153594
| 337
|
[
{
"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 with a cup in hand next to a horse. The horse begins to drink from the man's cup. Some kids and some dogs stand buy and watch. The man begins to extend his arm offering the cup to the horse. The horse proceeds to like the cup dry.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_IXUh06YCtjw.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_rAO-_VxIJng.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_EiXW33yuAcw.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_6q3EIv2X8BQ.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: "A person is seen playing an accordion on a busy street while many people walk around him and watch. The man continues to play the instrument and ends by stopping to laugh and smile at others."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_9GYLUAFgCXE.mp4",
"activitynet/videos/v_qNxLTF4Q6yk.mp4",
"activitynet/videos/v_3G9zc_SEOHM.mp4",
"activitynet/videos/v_X6CpfuJLx0U.mp4",
"activitynet/videos/v_2uBPhFis_4Y.mp4"
] | 4
|
<answer>4</answer>
|
[
4
] | true
| 1
| 24,411.233667
| 338
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames.\n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A person is seen playing an accordion on a busy street while many people walk around him and watch. The man continues to play the instrument and ends by stopping to laugh and smile at others.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_9GYLUAFgCXE.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_qNxLTF4Q6yk.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_3G9zc_SEOHM.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_X6CpfuJLx0U.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_2uBPhFis_4Y.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 guy helps girls onto a camel. The camel has a leaf in his nostril. The guy leads the camel with a rope attached to the camel face. A girl gets down from the camel. The girl walks underneath the camel."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_SokK_O2s9tQ.mp4",
"activitynet/videos/v_i49blayQ93Q.mp4",
"activitynet/videos/v_mi6f8kGVR70.mp4",
"activitynet/videos/v_xIld1Pt1QGs.mp4",
"activitynet/videos/v__jxpaVW4_cE.mp4"
] | 4
|
<answer>2</answer>
|
[
2
] | false
| 0
| 26,596.401546
| 340
|
[
{
"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 guy helps girls onto a camel. The camel has a leaf in his nostril. The guy leads the camel with a rope attached to the camel face. A girl gets down from the camel. The girl walks underneath the camel.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_SokK_O2s9tQ.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_i49blayQ93Q.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_mi6f8kGVR70.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_xIld1Pt1QGs.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__jxpaVW4_cE.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 ride horses in the shallow waters of the sea. People ride the horses in the water along the shore. A horse defecates in the water. People form a line with the horses while riding in the ocean."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_unE-vkRljRs.mp4",
"activitynet/videos/v_fid8KlncwTA.mp4",
"activitynet/videos/v_0EepbsAtiDk.mp4",
"activitynet/videos/v_hW6aZXhKl9M.mp4",
"activitynet/videos/v_M1-G6KEhY-M.mp4"
] | 4
|
<answer>4</answer>
|
[
4
] | true
| 1
| 36,212.435716
| 339
|
[
{
"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 ride horses in the shallow waters of the sea. People ride the horses in the water along the shore. A horse defecates in the water. People form a line with the horses while riding in the ocean.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_unE-vkRljRs.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_fid8KlncwTA.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_0EepbsAtiDk.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_hW6aZXhKl9M.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_M1-G6KEhY-M.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 throws a frisbee and the dog runs to get it. The man lays down and put his feet up, and the dog jumps on the feet. Then, the man throws the frisbee and the dog runs to get it. After, the man holds up the frisbee and the dogs stands in two legs. After, the dog pass between the legs of the person. Next, the person holds a frisbee high and the dog jumps to get the frisbee."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_Cx3QGeQu7xM.mp4",
"activitynet/videos/v_YAiCO8en_ls.mp4",
"activitynet/videos/v_vjMuhHo6wMY.mp4",
"activitynet/videos/v_iSIzuN9cEAs.mp4",
"activitynet/videos/v__7XW-BFK_ZY.mp4"
] | 1
|
<answer>1</answer>
|
[
1
] | true
| 1
| 29,268.182609
| 341
|
[
{
"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 throws a frisbee and the dog runs to get it. The man lays down and put his feet up, and the dog jumps on the feet. Then, the man throws the frisbee and the dog runs to get it. After, the man holds up the frisbee and the dogs stands in two legs. After, the dog pass between the legs of the person. Next, the person holds a frisbee high and the dog jumps to get the frisbee.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_Cx3QGeQu7xM.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_YAiCO8en_ls.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_vjMuhHo6wMY.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_iSIzuN9cEAs.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__7XW-BFK_ZY.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A large group of people are seen playing a soccer match on a large beach while the camera moves closer to watch them play. The kids continued to play up and down the beach and ends with pictures of the kids on the beach."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_5JlwYD_GChY.mp4",
"activitynet/videos/v_6j-H-tIjJvA.mp4",
"activitynet/videos/v_ncgzVLi_hlI.mp4",
"activitynet/videos/v_YeikEC85CGk.mp4",
"activitynet/videos/v_V6B8zFv1DdA.mp4"
] | 4
|
<answer>1</answer>
|
[
1
] | false
| 0
| 24,342.193284
| 342
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames.\n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A large group of people are seen playing a soccer match on a large beach while the camera moves closer to watch them play. The kids continued to play up and down the beach and ends with pictures of the kids on the beach.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_5JlwYD_GChY.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_6j-H-tIjJvA.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_ncgzVLi_hlI.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_YeikEC85CGk.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_V6B8zFv1DdA.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 front of a vanity mirror and brushes her hair. The woman stands up from a chair and her hair flows over the backrest. The long hair is tied up into a bun. A friend combs the woman's long hair. A friend braids the woman's hair while she sits."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_sdR443ncw-I.mp4",
"activitynet/videos/v_f1kY1-9XR1k.mp4",
"activitynet/videos/v_KNLGluuewIU.mp4",
"activitynet/videos/v_c_DQ7Y8ZRBQ.mp4",
"activitynet/videos/v_vw065HaGq3I.mp4"
] | 5
|
<answer>1</answer>
|
[
1
] | false
| 0
| 15,122.08522
| 343
|
[
{
"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 front of a vanity mirror and brushes her hair. The woman stands up from a chair and her hair flows over the backrest. The long hair is tied up into a bun. A friend combs the woman's long hair. A friend braids the woman's hair while she sits.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_sdR443ncw-I.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] 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
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_KNLGluuewIU.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_c_DQ7Y8ZRBQ.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_vw065HaGq3I.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 the ground id shown followed by a man stepping into frame and placing objects into a pile. The pile then catches flame and the man adds more to the fire to make it bigger."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_o8qR72Ymru8.mp4",
"activitynet/videos/v_HW9SFCj0dVU.mp4",
"activitynet/videos/v_CBoitanoH4s.mp4",
"activitynet/videos/v_LcXB-fSLTKY.mp4",
"activitynet/videos/v_Si6LZFiQT3k.mp4"
] | 3
|
<answer>2</answer>
|
[
2
] | false
| 0
| 17,623.052283
| 344
|
[
{
"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 the ground id shown followed by a man stepping into frame and placing objects into a pile. The pile then catches flame and the man adds more to the fire to make it bigger.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_o8qR72Ymru8.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_HW9SFCj0dVU.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_CBoitanoH4s.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_LcXB-fSLTKY.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_Si6LZFiQT3k.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 riding camels in a desert area. Two individuals that are leading the camels give each other high fives. A man and woman riding different camels give each other high fives."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_i49blayQ93Q.mp4",
"activitynet/videos/v_9xKOEE8Ni-Q.mp4",
"activitynet/videos/v_MkKUQ4MMHd8.mp4",
"activitynet/videos/v_dFsFL_WJasg.mp4",
"activitynet/videos/v_MmipoQF8EJs.mp4"
] | 3
|
<answer>2</answer>
|
[
2
] | false
| 0
| 11,423.800868
| 345
|
[
{
"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 riding camels in a desert area. Two individuals that are leading the camels give each other high fives. A man and woman riding different camels give each other high fives.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_i49blayQ93Q.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_9xKOEE8Ni-Q.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_MkKUQ4MMHd8.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_dFsFL_WJasg.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_MmipoQF8EJs.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A large group of people are standing around followed by two hugging and speaking to the camera. Several shot are shown of people playing beer pong as well as laughing with one another and yelling into the camera."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_gpKYclCmQHM.mp4",
"activitynet/videos/v_JDg--pjY5gg.mp4",
"activitynet/videos/v_EYoyxe8hd3g.mkv",
"activitynet/videos/v_J__1J4MmH4w.mp4",
"activitynet/videos/v_VtS4vy8Z0RQ.mp4"
] | 4
|
<answer>4</answer>
|
[
4
] | true
| 1
| 15,362.289629
| 346
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames.\n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A large group of people are standing around followed by two hugging and speaking to the camera. Several shot are shown of people playing beer pong as well as laughing with one another and yelling into 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_gpKYclCmQHM.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_JDg--pjY5gg.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_EYoyxe8hd3g.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_J__1J4MmH4w.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_VtS4vy8Z0RQ.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 for the clip is shown. A female styles her long, dark hair. Female tosses her hair and uses her fingers to play with her hair. The female displays a comb, clamp, and a scissor. The females uses the comb on her hair. The female parts her hair and puts some in a ponytail. The female places the clamp on her hair. The female cuts her hair after the clamp. The female removes the clamp. The female removes the ponytail band and uses her hands to style her hair. The credits for the video is shown."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_p1fpQ4yR1co.mp4",
"activitynet/videos/v_7O9kkDxEvaY.mp4",
"activitynet/videos/v_tOVv0cAyjcg.mp4",
"activitynet/videos/v_sMITf5WBIxM.mp4",
"activitynet/videos/v_dfjl7sS1IGo.mp4"
] | 3
|
<answer>2</answer>
|
[
2
] | false
| 0
| 16,300.209669
| 348
|
[
{
"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 for the clip is shown. A female styles her long, dark hair. Female tosses her hair and uses her fingers to play with her hair. The female displays a comb, clamp, and a scissor. The females uses the comb on her hair. The female parts her hair and puts some in a ponytail. The female places the clamp on her hair. The female cuts her hair after the clamp. The female removes the clamp. The female removes the ponytail band and uses her hands to style her hair. The credits for the video is shown.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_p1fpQ4yR1co.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_7O9kkDxEvaY.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_tOVv0cAyjcg.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_sMITf5WBIxM.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_dfjl7sS1IGo.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 men drive in a truck down a rural road. A man crawls through a field around elk. A large elk is seen running along a ridge. The elk gets shot and runs off. The elk falls over in pain onto the ground. The elk runs in slow motion being hit by an arrow. The hunter approaches the elk that is lying on the ground. The hunter examines the antlers of the elk."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_48zOi9j1E0A.mp4",
"activitynet/videos/v_3TNDCTlLlGk.mkv",
"activitynet/videos/v_EvJqfGXb5Fo.mp4",
"activitynet/videos/v_BFxxrjqgF0w.mp4",
"activitynet/videos/v_HCntSYltlmA.mp4"
] | 5
|
<answer>5</answer>
|
[
5
] | true
| 1
| 22,595.833145
| 347
|
[
{
"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 men drive in a truck down a rural road. A man crawls through a field around elk. A large elk is seen running along a ridge. The elk gets shot and runs off. The elk falls over in pain onto the ground. The elk runs in slow motion being hit by an arrow. The hunter approaches the elk that is lying on the ground. The hunter examines the antlers of the elk.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_48zOi9j1E0A.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_3TNDCTlLlGk.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_EvJqfGXb5Fo.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_BFxxrjqgF0w.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_HCntSYltlmA.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 a woman are on a dance floor and getting ready to dance. The man and woman begin dancing while people on the sides of the dance floor are watching them. When they're done dancing they depart and everyone watching begins clapping."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_ZUM89wyBcYY.mp4",
"activitynet/videos/v_MTJ1EtiizVQ.mp4",
"activitynet/videos/v_1MBVaveQDd8.mp4",
"activitynet/videos/v_lE3Hs4bsPhg.mkv",
"activitynet/videos/v_tm_CL7A0W4M.mp4"
] | 5
|
<answer>1</answer>
|
[
1
] | false
| 0
| 17,762.659062
| 349
|
[
{
"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 a woman are on a dance floor and getting ready to dance. The man and woman begin dancing while people on the sides of the dance floor are watching them. When they're done dancing they depart and everyone watching begins clapping.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_ZUM89wyBcYY.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_MTJ1EtiizVQ.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_1MBVaveQDd8.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_lE3Hs4bsPhg.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_tm_CL7A0W4M.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 sitting on a couch playing a harmonica. A man comes up in front of him and takes pictures of the man playing. The audience watching him is applauding."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_WpQHQeY43zo.mp4",
"activitynet/videos/v_cTioh2vzxGE.mp4",
"activitynet/videos/v_u02UsNRxclU.mp4",
"activitynet/videos/v_QhiKgeJV3k0.mkv",
"activitynet/videos/v_CQXhtaNkhrw.mp4"
] | 2
|
<answer>2</answer>
|
[
2
] | true
| 1
| 8,089.170161
| 350
|
[
{
"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 sitting on a couch playing a harmonica. A man comes up in front of him and takes pictures of the man playing. The audience watching him is applauding.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_WpQHQeY43zo.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_cTioh2vzxGE.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_u02UsNRxclU.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_QhiKgeJV3k0.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_CQXhtaNkhrw.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A gymnast mounts a low beam in front of the judges. He spins around in circles on the beam. He dismounts, his arms in the air."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_R0YS8JS_0rk.mp4",
"activitynet/videos/v_MVUqd8iVUEk.mp4",
"activitynet/videos/v_IJ76Wtgg2g4.mp4",
"activitynet/videos/v_3JHIcli-Wlg.mp4",
"activitynet/videos/v_dAcdSkaoK64.mp4"
] | 2
|
<answer>1</answer>
|
[
1
] | false
| 0
| 19,920.207249
| 352
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames.\n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A gymnast mounts a low beam in front of the judges. He spins around in circles on the beam. He dismounts, his arms in the air.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_R0YS8JS_0rk.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_MVUqd8iVUEk.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_IJ76Wtgg2g4.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_3JHIcli-Wlg.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_dAcdSkaoK64.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 shown putting plaster on the sides of a small, enclosed area. He holds an object in his hands and finally looks back towards the camera."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_Ny49eEt1OJg.mp4",
"activitynet/videos/v_rWQz-EwA4EA.mp4",
"activitynet/videos/v_E3UCEbGZmz0.mkv",
"activitynet/videos/v_rczR9C00KOM.mp4",
"activitynet/videos/v_o0gdMKlKLcU.mp4"
] | 2
|
<answer>4</answer>
|
[
4
] | false
| 0
| 20,950.429239
| 353
|
[
{
"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 shown putting plaster on the sides of a small, enclosed area. He holds an object in his hands and finally looks back towards 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_Ny49eEt1OJg.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_rWQz-EwA4EA.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_E3UCEbGZmz0.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_rczR9C00KOM.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_o0gdMKlKLcU.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 washing her hands in a sink and drying her hands off with a paper towel. The woman then waves her arms carelessly under the water and rubs her hands on her shirt to dry. She finally opens the door with holding a rag and then throws a rag away in the trash."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_tBNOJJx4Z9k.mp4",
"activitynet/videos/v_6g80a1NnftU.mp4",
"activitynet/videos/v_SAaqnGbci6Y.mp4",
"activitynet/videos/v_sS-KyhAzeUY.mkv",
"activitynet/videos/v_Lp3c3nwHrqM.mp4"
] | 2
|
<answer>2</answer>
|
[
2
] | true
| 1
| 28,399.562792
| 351
|
[
{
"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 washing her hands in a sink and drying her hands off with a paper towel. The woman then waves her arms carelessly under the water and rubs her hands on her shirt to dry. She finally opens the door with holding a rag and then throws a rag away in the trash.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_tBNOJJx4Z9k.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_6g80a1NnftU.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_SAaqnGbci6Y.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_sS-KyhAzeUY.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_Lp3c3nwHrqM.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 downstairs in what appears to be a basement turned into a gym is working out. He has a stepping mat in front of him and begins walking left to right going up and down on the step as he moves. With his hands by his waster moving back and forth,he continues on in the same manner with the exercise."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_Ule69iMpA3Y.mp4",
"activitynet/videos/v_8C1EFngZC3Q.mp4",
"activitynet/videos/v_OBDq689jDDY.mkv",
"activitynet/videos/v_puPMvwv2kmg.mp4",
"activitynet/videos/v_kbdBKIWKOWk.mkv"
] | 2
|
<answer>2</answer>
|
[
2
] | true
| 1
| 8,873.972527
| 354
|
[
{
"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 downstairs in what appears to be a basement turned into a gym is working out. He has a stepping mat in front of him and begins walking left to right going up and down on the step as he moves. With his hands by his waster moving back and forth,he continues on in the same manner with the exercise.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_Ule69iMpA3Y.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_8C1EFngZC3Q.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_OBDq689jDDY.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_puPMvwv2kmg.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_kbdBKIWKOWk.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A man is seen moving along the water on a surf board while another person watches on the side. The person continues riding around and slowing down to demonstrate how to play."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_jM3Buw2Kidk.mp4",
"activitynet/videos/v_L4rKeN_4CLk.mp4",
"activitynet/videos/v_CMGjxw3X1dI.mp4",
"activitynet/videos/v_qZk7okgCU2M.mp4",
"activitynet/videos/v_aoIGBV31OT4.mp4"
] | 2
|
<answer>2</answer>
|
[
2
] | true
| 1
| 24,309.776178
| 355
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames.\n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A man is seen moving along the water on a surf board while another person watches on the side. The person continues riding around and slowing down to demonstrate 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_jM3Buw2Kidk.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_L4rKeN_4CLk.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_CMGjxw3X1dI.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_qZk7okgCU2M.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_aoIGBV31OT4.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 image is blurry and clears to show men riding horses. We see the image multiples and see spectators. A lady in a white hat looks through binoculars at the game. Men are playing polo in a field. A girl in green runs behind the men with a flat and we see the lady in the hat with a video camera recording the event. We see he people playing polo again. We see the lady recording as she sits in her car and a player takes off his hat and sits in a chair before we see the game again."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_hCJTKVzkYFE.mp4",
"activitynet/videos/v_2_hcULoN4Ls.mp4",
"activitynet/videos/v_PUJYZEq8H64.mp4",
"activitynet/videos/v_fid8KlncwTA.mp4",
"activitynet/videos/v_rKtktLDSOpA.mp4"
] | 5
|
<answer>3</answer>
|
[
3
] | false
| 0
| 24,334.684832
| 357
|
[
{
"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 image is blurry and clears to show men riding horses. We see the image multiples and see spectators. A lady in a white hat looks through binoculars at the game. Men are playing polo in a field. A girl in green runs behind the men with a flat and we see the lady in the hat with a video camera recording the event. We see he people playing polo again. We see the lady recording as she sits in her car and a player takes off his hat and sits in a chair before we see the game again.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_hCJTKVzkYFE.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_2_hcULoN4Ls.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_PUJYZEq8H64.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_fid8KlncwTA.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_rKtktLDSOpA.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 standing flying kites with another young man on a street and they are really happy. They twirl the strings around in the air. The boy twirls the string hard around to make the kite fly while the girls continue as well. The kites are high in the sky and a woman walks by and watches as another woman walks towards the three kids. Another young boy in a blue shirt walks by the crowd and leans and watches the three."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_NGk3v4sKqdg.mp4",
"activitynet/videos/v_IqRN2sOQ7Mo.mp4",
"activitynet/videos/v_1B3XsffrM4M.mp4",
"activitynet/videos/v_TqcoukXhXeA.mp4",
"activitynet/videos/v_pTmlOZY0e4c.mp4"
] | 4
|
<answer>2</answer>
|
[
2
] | false
| 0
| 28,050.825039
| 356
|
[
{
"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 standing flying kites with another young man on a street and they are really happy. They twirl the strings around in the air. The boy twirls the string hard around to make the kite fly while the girls continue as well. The kites are high in the sky and a woman walks by and watches as another woman walks towards the three kids. Another young boy in a blue shirt walks by the crowd and leans and watches the three.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_NGk3v4sKqdg.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_IqRN2sOQ7Mo.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_1B3XsffrM4M.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_TqcoukXhXeA.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_pTmlOZY0e4c.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 demonstrates the safe use of a bissell steam cleaner on a hardwood floor. The woman begins by talking, looking into the camera and standing next to, and holding the handle of a steam cleaner while standing on a hardwood floor and pulling the cleaner across the floor with white steam pouring out of it. The steam cleaner then stands on its own with closeups of its hardware, and buttons before the woman begins pulling the cleaner across the floor again. Lastly, the woman stands beside the cleaner again, talks, and makes one last walk forward with the cleaner and a closeup of the woman's ballet flats."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v__MR8G1jwM4o.mp4",
"activitynet/videos/v_kWmf0_XSfBU.mp4",
"activitynet/videos/v_hc4DBHpRuGM.mp4",
"activitynet/videos/v_6ChRD-1NwSg.mp4",
"activitynet/videos/v_7Eh6c1eYMFk.mkv"
] | 4
|
<answer>4</answer>
|
[
4
] | true
| 1
| 13,998.416199
| 358
|
[
{
"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 demonstrates the safe use of a bissell steam cleaner on a hardwood floor. The woman begins by talking, looking into the camera and standing next to, and holding the handle of a steam cleaner while standing on a hardwood floor and pulling the cleaner across the floor with white steam pouring out of it. The steam cleaner then stands on its own with closeups of its hardware, and buttons before the woman begins pulling the cleaner across the floor again. Lastly, the woman stands beside the cleaner again, talks, and makes one last walk forward with the cleaner and a closeup of the woman's ballet flats.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v__MR8G1jwM4o.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_kWmf0_XSfBU.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_hc4DBHpRuGM.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_6ChRD-1NwSg.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_7Eh6c1eYMFk.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 and leads into clips of people speaking to one another. More clips are shown of people riding on snowboards and speaking to one another. A person is then seen riding behind a car on a snowboard while others drive up to speak to him and record his time."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_a1WhnMcTbrY.mp4",
"activitynet/videos/v_wpxozv4Yois.mp4",
"activitynet/videos/v_PAiJNr97C6g.mp4",
"activitynet/videos/v_kXbc9D0sF5k.mp4",
"activitynet/videos/v_dUzqM38vwPI.mp4"
] | 5
|
<answer>2</answer>
|
[
2
] | false
| 0
| 25,197.286369
| 359
|
[
{
"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 and leads into clips of people speaking to one another. More clips are shown of people riding on snowboards and speaking to one another. A person is then seen riding behind a car on a snowboard while others drive up to speak to him and record his time.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_a1WhnMcTbrY.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_wpxozv4Yois.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_PAiJNr97C6g.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_kXbc9D0sF5k.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_dUzqM38vwPI.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 teams are gathered on an outdoor field. They are running and trying to hit a ball with a bat. They try to win the game, hitting the ball toward the opponent's goal."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_YVxuIAwOyoE.mp4",
"activitynet/videos/v_pHiulmPx7ek.mp4",
"activitynet/videos/v_BSdXxBOJ12A.mp4",
"activitynet/videos/v_ew7XlNRrKyM.mp4",
"activitynet/videos/v_DRSH-_Ye9eE.mp4"
] | 4
|
<answer>2</answer>
|
[
2
] | false
| 0
| 29,774.744744
| 360
|
[
{
"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 teams are gathered on an outdoor field. They are running and trying to hit a ball with a bat. They try to win the game, hitting the ball toward the opponent's goal.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_YVxuIAwOyoE.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_pHiulmPx7ek.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_BSdXxBOJ12A.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_ew7XlNRrKyM.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_DRSH-_Ye9eE.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 teen puts his left leg on the legs of a youth. Then, the youth peels wax paper and put on the leg of the teen. Next, the youth peels off the wax paper and shows the hair in the paper. The teen shows his leg making pain gestures on his face."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_0fsMeZoZzJI.mp4",
"activitynet/videos/v_gh0GD6OvLHE.mp4",
"activitynet/videos/v_PLnfT1PoVHw.mp4",
"activitynet/videos/v_HtkuvF7VbSQ.mp4",
"activitynet/videos/v_gFv3PrFkeL0.mp4"
] | 3
|
<answer>2</answer>
|
[
2
] | false
| 0
| 9,792.99004
| 362
|
[
{
"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 teen puts his left leg on the legs of a youth. Then, the youth peels wax paper and put on the leg of the teen. Next, the youth peels off the wax paper and shows the hair in the paper. The teen shows his leg making pain gestures on his face.\"\n\nCandidates:\n[1] 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[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_PLnfT1PoVHw.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_HtkuvF7VbSQ.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_gFv3PrFkeL0.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 seen pushing a lawn mower along the side of a lawn while looking and laughing the camera. The man turns around and continues mowing along the grass while the camera captures his movements."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_9wRQsxVFwkE.mp4",
"activitynet/videos/v_iHO42zwYsu0.mp4",
"activitynet/videos/v_62BPME-ikJU.mp4",
"activitynet/videos/v_2MRR5NxbO9k.mp4",
"activitynet/videos/v_CGIrDfEP5lE.mp4"
] | 1
|
<answer>4</answer>
|
[
4
] | false
| 0
| 26,171.802976
| 361
|
[
{
"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 seen pushing a lawn mower along the side of a lawn while looking and laughing the camera. The man turns around and continues mowing along the grass while the camera captures his movements.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_9wRQsxVFwkE.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_iHO42zwYsu0.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_62BPME-ikJU.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[4] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_2MRR5NxbO9k.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_CGIrDfEP5lE.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 woman is putting on makeup and looking into a mirror. The woman continues putting makeup on as the camera watches her from the side."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_UPfQNZl0_dg.mp4",
"activitynet/videos/v_tScqYRQ7zyo.mp4",
"activitynet/videos/v_-uR5-jYe0Ag.mp4",
"activitynet/videos/v_dZZE8HI0OBE.mp4",
"activitynet/videos/v_Pcro3S-4EnQ.mkv"
] | 2
|
<answer>2</answer>
|
[
2
] | true
| 1
| 7,163.162486
| 363
|
[
{
"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 woman is putting on makeup and looking into a mirror. The woman continues putting makeup on as the camera watches her from the side.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_UPfQNZl0_dg.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_tScqYRQ7zyo.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_-uR5-jYe0Ag.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_dZZE8HI0OBE.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_Pcro3S-4EnQ.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 animals are seen on the ground followed by a cat and dog. A cat is shown carrying a leash in it's mouth with a dog following close behind. The camera follows the two walk up a set of stairs."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_ROMy00dG8Ds.mp4",
"activitynet/videos/v_ak5mpw8komA.mp4",
"activitynet/videos/v_yPA6klGWEsc.mp4",
"activitynet/videos/v_Fky1ioAUt38.mp4",
"activitynet/videos/v_NQOPahBcpSE.mp4"
] | 2
|
<answer>2</answer>
|
[
2
] | true
| 1
| 25,819.086453
| 365
|
[
{
"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 animals are seen on the ground followed by a cat and dog. A cat is shown carrying a leash in it's mouth with a dog following close behind. The camera follows the two walk up a set of stairs.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_ROMy00dG8Ds.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_ak5mpw8komA.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_yPA6klGWEsc.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_Fky1ioAUt38.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_NQOPahBcpSE.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: "Chocolate chip cookies appear on a square white plate and a hand grabs one, bites it, and dips it into a cup of milk. The person is now showing a very close up of them putting ingredients into a large clear glass bowl, and they mix it all together with a wooden spoon. The person stops mixing with a spoon and then begins to use their hand to mix it all together. A package of chocolates are shown, then mixed into the cookie dough by hand, then covered with saran wrap. After a while the saran wrap is removed and small coops of the dough is put onto a cooking mat, baked in the oven, and cooled. The cookies are now on the white plate and the person takes a bite of one and also dips it in milk. The outro appears and it's of various different creations and ends with a screen that has 4 small screens on it with the words above it that say "Click below for more videos!"."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_Uqs8NaPzHKU.mp4",
"activitynet/videos/v_xsdrqauYhJs.mp4",
"activitynet/videos/v_VaT3qsoHPQ8.mp4",
"activitynet/videos/v_ZTHsS5lQyvQ.mp4",
"activitynet/videos/v_F3jJVS3NHf8.mp4"
] | 3
|
<answer>3</answer>
|
[
3
] | true
| 1
| 30,410.640515
| 366
|
[
{
"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: \"Chocolate chip cookies appear on a square white plate and a hand grabs one, bites it, and dips it into a cup of milk. The person is now showing a very close up of them putting ingredients into a large clear glass bowl, and they mix it all together with a wooden spoon. The person stops mixing with a spoon and then begins to use their hand to mix it all together. A package of chocolates are shown, then mixed into the cookie dough by hand, then covered with saran wrap. After a while the saran wrap is removed and small coops of the dough is put onto a cooking mat, baked in the oven, and cooled. The cookies are now on the white plate and the person takes a bite of one and also dips it in milk. The outro appears and it's of various different creations and ends with a screen that has 4 small screens on it with the words above it that say \"Click below for more videos!\".\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_Uqs8NaPzHKU.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_xsdrqauYhJs.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_VaT3qsoHPQ8.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_ZTHsS5lQyvQ.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_F3jJVS3NHf8.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 video leads into several shots of people throwing bowling balls down a lane and hitting numerous pins. The video continues with more shots of people throwing bowling balls and ending with a title screen."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_y56qXoJh6U0.mp4",
"activitynet/videos/v_w8kVVzMOC98.mp4",
"activitynet/videos/v_6gyD-Mte2ZM.mp4",
"activitynet/videos/v_Jd0KWW9LN4Q.mp4",
"activitynet/videos/v_5GiIqXY__74.mp4"
] | 5
|
<answer>1</answer>
|
[
1
] | false
| 0
| 37,808.76001
| 364
|
[
{
"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 video leads into several shots of people throwing bowling balls down a lane and hitting numerous pins. The video continues with more shots of people throwing bowling balls and ending with a 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_y56qXoJh6U0.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_w8kVVzMOC98.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_6gyD-Mte2ZM.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_Jd0KWW9LN4Q.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_5GiIqXY__74.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 swings a ball on a chain around his body and throws it behind him. A man in a blue shirt stands behind a fence behind him."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_8gUKEh27AFM.mp4",
"activitynet/videos/v_5g7bqiT7Y3c.mp4",
"activitynet/videos/v_HtCQ-OmHJl4.mp4",
"activitynet/videos/v_NbO4k5EtU4Y.mp4",
"activitynet/videos/v_oIEDMaMo7UE.mp4"
] | 1
|
<answer>1</answer>
|
[
1
] | true
| 1
| 18,512.823566
| 367
|
[
{
"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 swings a ball on a chain around his body and throws it behind him. A man in a blue shirt stands behind a fence behind him.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_8gUKEh27AFM.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_5g7bqiT7Y3c.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_HtCQ-OmHJl4.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_NbO4k5EtU4Y.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_oIEDMaMo7UE.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 hosting a news segment and leads into clips of kids playing in a pool. Parents are then seen walking around and putting sunscreen on their kids and leads back into the woman hosting."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_1uiEkwykOxo.mp4",
"activitynet/videos/v_7S15OsGinjw.mp4",
"activitynet/videos/v_mf6UsZuW9Nw.mp4",
"activitynet/videos/v_kGvs0Nv5zJo.mp4",
"activitynet/videos/v_5TMKHLOACYg.mp4"
] | 3
|
<answer>2</answer>
|
[
2
] | false
| 0
| 12,257.462456
| 368
|
[
{
"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 hosting a news segment and leads into clips of kids playing in a pool. Parents are then seen walking around and putting sunscreen on their kids and leads back into the woman hosting.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_1uiEkwykOxo.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_7S15OsGinjw.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_mf6UsZuW9Nw.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_kGvs0Nv5zJo.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_5TMKHLOACYg.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 four people are playing volleyball outdoors. They are shown in slow motion as they lob the ball back and forth. The speed increases as the game continues."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_Uc1_7BXtXZs.mp4",
"activitynet/videos/v_4_wfCFTnExI.mp4",
"activitynet/videos/v_kQ4rE7o6rrg.mp4",
"activitynet/videos/v_-MB6Wxglgzw.mp4",
"activitynet/videos/v_s_QH-5G33Fw.mp4"
] | 2
|
<answer>2</answer>
|
[
2
] | true
| 1
| 17,385.483676
| 369
|
[
{
"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 four people are playing volleyball outdoors. They are shown in slow motion as they lob the ball back and forth. The speed increases as the game continues.\"\n\nCandidates:\n[1] 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
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_4_wfCFTnExI.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_kQ4rE7o6rrg.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_-MB6Wxglgzw.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_s_QH-5G33Fw.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "The woman is standing at the counter. The woman is showing an two side metal shot glass, then she pour vodka in the cup and pour in the tall glass. The woman is holding a bottle of vodka and the metal shot glass, pour the liquid in the cup and pour the liquid in the pot. The woman pour the liquid in a shot glass, then pour it in the cup. The woman put ice in the glass and pour vodka."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_vvvjTjsXbzE.mkv",
"activitynet/videos/v_uMGfCaGMnEE.mp4",
"activitynet/videos/v_20ooSJixdyg.mp4",
"activitynet/videos/v_03JdaRepHkA.mp4",
"activitynet/videos/v_hoYF0DhYVOI.mp4"
] | 1
|
<answer>4</answer>
|
[
4
] | false
| 0
| 15,917.459966
| 370
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames.\n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"The woman is standing at the counter. The woman is showing an two side metal shot glass, then she pour vodka in the cup and pour in the tall glass. The woman is holding a bottle of vodka and the metal shot glass, pour the liquid in the cup and pour the liquid in the pot. The woman pour the liquid in a shot glass, then pour it in the cup. The woman put ice in the glass and pour vodka.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_vvvjTjsXbzE.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_uMGfCaGMnEE.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_20ooSJixdyg.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_03JdaRepHkA.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_hoYF0DhYVOI.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 prepares for a high jump. The man then completes the high jump successfully."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_nLAm7USuYbA.mp4",
"activitynet/videos/v_Cgquef_qgcs.mp4",
"activitynet/videos/v_IqXaLlFSWwc.mp4",
"activitynet/videos/v_bZF4nakRNF4.mp4",
"activitynet/videos/v_Z47QGlaQ1NE.mp4"
] | 3
|
<answer>1</answer>
|
[
1
] | false
| 0
| 20,576.169057
| 371
|
[
{
"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 prepares for a high jump. The man then completes the high jump successfully.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_nLAm7USuYbA.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_Cgquef_qgcs.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_IqXaLlFSWwc.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_bZF4nakRNF4.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_Z47QGlaQ1NE.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "Two men stand and discuss in a gym with weight racks. The weight lifter squats down to show proper stance when lifting. The trainer presses on the weightlifter's back to show correct posture. The weightlifter stands back up after the presentation."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_NpBZn7OHUKo.mp4",
"activitynet/videos/v_zh0haUMeZV0.mp4",
"activitynet/videos/v_DV1ITGBfo5w.mp4",
"activitynet/videos/v_W8XwSNt8P5A.mp4",
"activitynet/videos/v__Zq8ugolzlA.mp4"
] | 5
|
<answer>2</answer>
|
[
2
] | false
| 0
| 18,603.684208
| 372
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames.\n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"Two men stand and discuss in a gym with weight racks. The weight lifter squats down to show proper stance when lifting. The trainer presses on the weightlifter's back to show correct posture. The weightlifter stands back up after the presentation.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_NpBZn7OHUKo.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_zh0haUMeZV0.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_DV1ITGBfo5w.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_W8XwSNt8P5A.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__Zq8ugolzlA.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 a camera and leads into video surveillance of two men walking into a store then driving away. Shots are shown of the store, cigarettes, and people smoking cigarettes, followed by another man being interviewed and smiling to the camera. One last man is seen holding a pack of cigarettes and speaking to the camera."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_ulopyhvgyQg.mp4",
"activitynet/videos/v_aBdrTqSnWbw.mp4",
"activitynet/videos/v_R8WbSI3m1lI.mp4",
"activitynet/videos/v_e4AIrRnWakA.mp4",
"activitynet/videos/v_DfOiHMcrCbs.mp4"
] | 5
|
<answer>5</answer>
|
[
5
] | true
| 1
| 12,065.386159
| 374
|
[
{
"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 a camera and leads into video surveillance of two men walking into a store then driving away. Shots are shown of the store, cigarettes, and people smoking cigarettes, followed by another man being interviewed and smiling to the camera. One last man is seen holding a pack of cigarettes and speaking 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_ulopyhvgyQg.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_aBdrTqSnWbw.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_R8WbSI3m1lI.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_e4AIrRnWakA.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_DfOiHMcrCbs.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 landscape in the desert. We see a man talking. We then see one man climbing a sheer cliff. We see the climber from a distance. We see the man from a distance and as he reaches the top. We see the man talking and hug his dog. We then see the ending screen again."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_6Yn2U58qxPs.mp4",
"activitynet/videos/v_Pr6zL1ToSC4.mp4",
"activitynet/videos/v_UALnEw4XhTY.mp4",
"activitynet/videos/v_eYgXvnnlPQA.mp4",
"activitynet/videos/v__4LZrf1GL1s.mp4"
] | 4
|
<answer>4</answer>
|
[
4
] | true
| 1
| 30,179.261467
| 373
|
[
{
"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 landscape in the desert. We see a man talking. We then see one man climbing a sheer cliff. We see the climber from a distance. We see the man from a distance and as he reaches the top. We see the man talking and hug his dog. We then see the ending screen again.\"\n\nCandidates:\n[1] 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[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_Pr6zL1ToSC4.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_UALnEw4XhTY.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_eYgXvnnlPQA.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__4LZrf1GL1s.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 poker table is seen with people sitting around the sides and speaking to one anther. The man is seen playing with the others and putting his cars in the middle."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_jimvzigX1ak.mp4",
"activitynet/videos/v_F30odTEdsxo.mp4",
"activitynet/videos/v_bc7r5_gSAVg.mp4",
"activitynet/videos/v_UGujWA07GkM.mp4",
"activitynet/videos/v_LCLDhKiMAPA.mp4"
] | 2
|
<answer>4</answer>
|
[
4
] | false
| 0
| 22,419.897087
| 375
|
[
{
"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 poker table is seen with people sitting around the sides and speaking to one anther. The man is seen playing with the others and putting his cars in the middle.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_jimvzigX1ak.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_F30odTEdsxo.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_bc7r5_gSAVg.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_UGujWA07GkM.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_LCLDhKiMAPA.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 cesar salad is eady and its served in a bowl. croutons are in a bowl and chopped ingredients are separated. theman mix all the ingredients in a bowl to make the dressing, put plastic wrap as a lid. man cutts the lettuce and in a pan put oil with garlic and stir fry the croutons. the man put the dressing on the lettuces and add the croutons in the bowl and mix them all together."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_B5hzlU0OepQ.mp4",
"activitynet/videos/v_sf2zGT5nN04.mp4",
"activitynet/videos/v_LoBjzA2z2Ls.mp4",
"activitynet/videos/v_t97xM9sY2yg.mp4",
"activitynet/videos/v_3X6eP273RoI.mp4"
] | 3
|
<answer>1</answer>
|
[
1
] | false
| 0
| 26,631.57703
| 376
|
[
{
"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 cesar salad is eady and its served in a bowl. croutons are in a bowl and chopped ingredients are separated. theman mix all the ingredients in a bowl to make the dressing, put plastic wrap as a lid. man cutts the lettuce and in a pan put oil with garlic and stir fry the croutons. the man put the dressing on the lettuces and add the croutons in the bowl and mix them all together.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_B5hzlU0OepQ.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_sf2zGT5nN04.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_LoBjzA2z2Ls.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_t97xM9sY2yg.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_3X6eP273RoI.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 shows the proper way to do crunches on a pier near the ocean. She says not to straigh your neck and not to lift your head with arms."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_PwOMgya8qYI.mp4",
"activitynet/videos/v_7tlXgKBTD_0.mp4",
"activitynet/videos/v_A_ndiCY-rDc.mp4",
"activitynet/videos/v_hYRNSJwhVPw.mkv",
"activitynet/videos/v_5vd8j0hKIgs.mp4"
] | 4
|
<answer>4</answer>
|
[
4
] | true
| 1
| 15,926.912371
| 377
|
[
{
"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 shows the proper way to do crunches on a pier near the ocean. She says not to straigh your neck and not to lift your head with arms.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_PwOMgya8qYI.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_7tlXgKBTD_0.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_A_ndiCY-rDc.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_hYRNSJwhVPw.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_5vd8j0hKIgs.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A large group of people are seen standing around a gymnasium that leads into several clips of people throwing balls to one another. More clips are shown of people kicking balls into goals and others blocking them. The players continue practicing drills with one another and ends with text going up the screen."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_RihO8i98QJg.mp4",
"activitynet/videos/v_l_KhWbeZeRA.mkv",
"activitynet/videos/v_XIMi2oydVB8.mp4",
"activitynet/videos/v_FMYu8k1b_DM.mp4",
"activitynet/videos/v_IoGpS8NQklE.mp4"
] | 4
|
<answer>1</answer>
|
[
1
] | false
| 0
| 22,088.006026
| 378
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames.\n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A large group of people are seen standing around a gymnasium that leads into several clips of people throwing balls to one another. More clips are shown of people kicking balls into goals and others blocking them. The players continue practicing drills with one another and ends with text going up 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_RihO8i98QJg.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_l_KhWbeZeRA.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_XIMi2oydVB8.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_FMYu8k1b_DM.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_IoGpS8NQklE.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A little girl hits a minion character in front of a group of children and adults. Everyone cheers and there is some clapping. The little girl turns around and returns her stick to the woman."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_NURr5XJcwFs.mp4",
"activitynet/videos/v_-GRvxWH4axc.mp4",
"activitynet/videos/v_VWpRBfhoFVg.mp4",
"activitynet/videos/v_w7IeqGuuA7Q.mp4",
"activitynet/videos/v_N6ERAg1EKcc.mp4"
] | 3
|
<answer>2</answer>
|
[
2
] | false
| 0
| 16,236.936039
| 380
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames.\n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A little girl hits a minion character in front of a group of children and adults. Everyone cheers and there is some clapping. The little girl turns around and returns her stick to the woman.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_NURr5XJcwFs.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_-GRvxWH4axc.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_VWpRBfhoFVg.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_w7IeqGuuA7Q.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_N6ERAg1EKcc.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 in rooms playing wall ball. A woman in a blue shirt stands outside in the hallway."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_9FK7tjzBKio.mp4",
"activitynet/videos/v_TotbMcWIoyE.mp4",
"activitynet/videos/v_4mlA78hn4mY.mp4",
"activitynet/videos/v_qcsGJTJstZ4.mp4",
"activitynet/videos/v_Re-SsHmajds.mp4"
] | 1
|
<answer>5</answer>
|
[
5
] | false
| 0
| 22,153.190654
| 379
|
[
{
"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 in rooms playing wall ball. A woman in a blue shirt stands outside in the hallway.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_9FK7tjzBKio.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_TotbMcWIoyE.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_4mlA78hn4mY.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_qcsGJTJstZ4.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_Re-SsHmajds.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 kneling down in the floor paint a white cloth in a living room. man is walking by a halfway and puth the courtains in the wall by the window."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_doZb3RlLSts.mp4",
"activitynet/videos/v_GLL1vOrV5Qo.mp4",
"activitynet/videos/v_P2Fcv3cC8bI.mp4",
"activitynet/videos/v_Dk3DiAp2yAU.mp4",
"activitynet/videos/v_-U4lNtzVQ8s.mp4"
] | 3
|
<answer>1</answer>
|
[
1
] | false
| 0
| 9,608.217005
| 383
|
[
{
"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 kneling down in the floor paint a white cloth in a living room. man is walking by a halfway and puth the courtains in the wall by the window.\"\n\nCandidates:\n[1] 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[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_GLL1vOrV5Qo.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_P2Fcv3cC8bI.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_Dk3DiAp2yAU.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_-U4lNtzVQ8s.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 shown in several clips running down a track, jumping into a pit of sand. He pauses for a woman and then continues jumping back into the pit several more times."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_AJ15GW-sS5M.mp4",
"activitynet/videos/v_bZ4r3Y_qceE.mp4",
"activitynet/videos/v_rdkrg8Bj9_I.mp4",
"activitynet/videos/v_E2nAOID5DLM.mp4",
"activitynet/videos/v_VvD2fdPNWEg.mp4"
] | 3
|
<answer>2</answer>
|
[
2
] | false
| 0
| 25,678.659836
| 382
|
[
{
"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 shown in several clips running down a track, jumping into a pit of sand. He pauses for a woman and then continues jumping back into the pit several more times.\"\n\nCandidates:\n[1] 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
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_bZ4r3Y_qceE.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_rdkrg8Bj9_I.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_E2nAOID5DLM.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_VvD2fdPNWEg.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 on the right serves the ball. The two men hit the ball back and forth. The right man misses the ball and we see the people on the sidelines. The right man serves again. They go back and forth till the ball drops again 1:10. 1:14 we see the crowd before the left man serves and the go back and forth. 124 The right man misses, the left throws his hands in the air and walks off."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_uz04njTFKP8.mp4",
"activitynet/videos/v_cyXWvxVt8qE.mp4",
"activitynet/videos/v_A32TgJfp2z8.mp4",
"activitynet/videos/v_xRuZMDClaQM.mp4",
"activitynet/videos/v_evkiciK1nRc.mp4"
] | 2
|
<answer>1</answer>
|
[
1
] | false
| 0
| 33,156.550627
| 381
|
[
{
"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 on the right serves the ball. The two men hit the ball back and forth. The right man misses the ball and we see the people on the sidelines. The right man serves again. They go back and forth till the ball drops again 1:10. 1:14 we see the crowd before the left man serves and the go back and forth. 124 The right man misses, the left throws his hands in the air and walks off.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_uz04njTFKP8.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_cyXWvxVt8qE.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_A32TgJfp2z8.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_xRuZMDClaQM.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_evkiciK1nRc.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "Men in jerseys bent down on the field trying to figure out a play. They are conversing about what the next move is going to be. The coaches are standing by talking to the player explaining. Mean while two players in the back play rock paper scissors."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_YZvdzvM-124.mp4",
"activitynet/videos/v_agM7yjqVKo0.mp4",
"activitynet/videos/v_NhcOmldkGIo.mp4",
"activitynet/videos/v_7X3wPRKuAsU.mp4",
"activitynet/videos/v_x86YIU9TIPw.mp4"
] | 1
|
<answer>5</answer>
|
[
5
] | false
| 0
| 15,422.815588
| 384
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames.\n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"Men in jerseys bent down on the field trying to figure out a play. They are conversing about what the next move is going to be. The coaches are standing by talking to the player explaining. Mean while two players in the back play rock paper scissors.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_YZvdzvM-124.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_agM7yjqVKo0.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_NhcOmldkGIo.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_7X3wPRKuAsU.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_x86YIU9TIPw.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: "These women are playing volleyball in the gymnasium. Some of these girls are wearing a white and blue outfit and the other team is wearing black and gold outfits."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_E3dV8LdAPx8.mp4",
"activitynet/videos/v_s56ctLdnOdw.mkv",
"activitynet/videos/v_ke3R1rOeQzE.mkv",
"activitynet/videos/v_laeOL4ipHck.mp4",
"activitynet/videos/v_aYtnkEWM_Cg.mp4"
] | 2
|
<answer>4</answer>
|
[
4
] | false
| 0
| 11,591.066883
| 385
|
[
{
"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: \"These women are playing volleyball in the gymnasium. Some of these girls are wearing a white and blue outfit and the other team is wearing black and gold outfits.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_E3dV8LdAPx8.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_s56ctLdnOdw.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_ke3R1rOeQzE.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_aYtnkEWM_Cg.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 monkey is seen walking up onto a tree branch. The camera pans around and back to the monkey who is now sitting on the branch. The monkey is then seen eating food and looking at the camera."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_0HhNhRExwSQ.mp4",
"activitynet/videos/v_x4c_wI6kQyE.mp4",
"activitynet/videos/v_jHXqbgeq83Y.mp4",
"activitynet/videos/v_WaFDgdqY1DM.mp4",
"activitynet/videos/v_dzR4voNDZ7g.mp4"
] | 4
|
<answer>4</answer>
|
[
4
] | true
| 1
| 12,350.58142
| 386
|
[
{
"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 monkey is seen walking up onto a tree branch. The camera pans around and back to the monkey who is now sitting on the branch. The monkey is then seen eating food and looking at the camera.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_0HhNhRExwSQ.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_x4c_wI6kQyE.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_jHXqbgeq83Y.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_WaFDgdqY1DM.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_dzR4voNDZ7g.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 wearing helmets and safety vests pull a raft into a rapid river. They begin white water rafting. They run through the water toward a water fall, celebrating. They then pick up their raft and carrying it back to their van before loading up to leave."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_jQHGyqk21GI.mp4",
"activitynet/videos/v_f6wAW0Jv2Eo.mkv",
"activitynet/videos/v_-sd2XAFkeC0.mkv",
"activitynet/videos/v_jzNdWjZm92s.mp4",
"activitynet/videos/v_gKLbdLKEG6U.mp4"
] | 5
|
<answer>2</answer>
|
[
2
] | false
| 0
| 27,221.404066
| 387
|
[
{
"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 wearing helmets and safety vests pull a raft into a rapid river. They begin white water rafting. They run through the water toward a water fall, celebrating. They then pick up their raft and carrying it back to their van before loading up to leave.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_jQHGyqk21GI.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_f6wAW0Jv2Eo.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_-sd2XAFkeC0.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_jzNdWjZm92s.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_gKLbdLKEG6U.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "We see two young people playing shuffle board on the deck of a cruise ship. We see a speedboat riding on a the ocean. We switch to two adults playing shuffleboard. The camera turns on it's side then straight. The little girl holding the camera speaks to the camera. The little girl turns the camera on herself and talks again."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_aFSaGCvYXXY.mkv",
"activitynet/videos/v_wDFpFJ1CP9g.mkv",
"activitynet/videos/v_RL4V-Sx619M.mp4",
"activitynet/videos/v_XuG2V9gDD9M.mp4",
"activitynet/videos/v_hmPeCPjaxAM.mp4"
] | 3
|
<answer>3</answer>
|
[
3
] | true
| 1
| 33,641.376089
| 388
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames.\n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"We see two young people playing shuffle board on the deck of a cruise ship. We see a speedboat riding on a the ocean. We switch to two adults playing shuffleboard. The camera turns on it's side then straight. The little girl holding the camera speaks to the camera. The little girl turns the camera on herself and talks again.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_aFSaGCvYXXY.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_wDFpFJ1CP9g.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_RL4V-Sx619M.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_XuG2V9gDD9M.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_hmPeCPjaxAM.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 screen appears with a gray background and the red and white words that say PRO TRAINING. A man appears and his twitter name handle appears on the bottom left while he's talking in a basketball court. A black screen appears and the white words THE MIKAN DRILL are displayed and the man continues to talk and then demonstrates exactly how to do various Mikan Drills. The man then grabs a second ball and continues to demonstrate advanced Mikan Drills with both balls. When the man is done with the drills he's talking again and the outro appears with his contact information and a subscribe button while a small video shows some of his other skills."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_gMLA3a0FMS0.mp4",
"activitynet/videos/v_ZlwkO1oFBHw.mp4",
"activitynet/videos/v_UfjR8ewF8xo.mp4",
"activitynet/videos/v_ssBiSN8XofA.mp4",
"activitynet/videos/v_MSPslSgkp60.mp4"
] | 4
|
<answer>4</answer>
|
[
4
] | true
| 1
| 38,590.078531
| 389
|
[
{
"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 screen appears with a gray background and the red and white words that say PRO TRAINING. A man appears and his twitter name handle appears on the bottom left while he's talking in a basketball court. A black screen appears and the white words THE MIKAN DRILL are displayed and the man continues to talk and then demonstrates exactly how to do various Mikan Drills. The man then grabs a second ball and continues to demonstrate advanced Mikan Drills with both balls. When the man is done with the drills he's talking again and the outro appears with his contact information and a subscribe button while a small video shows some of his other skills.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_gMLA3a0FMS0.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_ZlwkO1oFBHw.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_ssBiSN8XofA.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_MSPslSgkp60.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 holding a contact lens in her hand. She puts the contact lens in her eyes. She turns around and smiles."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_Uj1QtIM8500.mp4",
"activitynet/videos/v_l5QQ1vVctOo.mp4",
"activitynet/videos/v_jxk6KOLu5kU.mp4",
"activitynet/videos/v_qVuRcevXgMk.mp4",
"activitynet/videos/v_HsGz6S2MBU4.mp4"
] | 4
|
<answer>3</answer>
|
[
3
] | false
| 0
| 28,059.956333
| 390
|
[
{
"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 holding a contact lens in her hand. She puts the contact lens in her eyes. She turns around and smiles.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_Uj1QtIM8500.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_l5QQ1vVctOo.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_jxk6KOLu5kU.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_qVuRcevXgMk.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_HsGz6S2MBU4.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 kids are outside of their school and they jump up with joy and the title of their school is then shown. A kid then takes a rock and throws it at a number on the hopscotch drawn on the ground with chalk. The video then begins by showing the steps to hopscotch and a boy begins drawing boxes and adding numbers up to the number 8. After, a marker is shown which happens to be a box and they throw it on the number 1 and then proceed to hop over the number one and come back and pick up the rock. The game continues and the line of boys throw the rock and take turns until the game is over."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_iDMzTPfELoc.mp4",
"activitynet/videos/v_hIJ6VTEKji8.mp4",
"activitynet/videos/v_8Kv7piYg9yc.mp4",
"activitynet/videos/v_lKKimizxQJM.mp4",
"activitynet/videos/v_msiX-xky6Ac.mp4"
] | 5
|
<answer>5</answer>
|
[
5
] | true
| 1
| 15,749.990054
| 391
|
[
{
"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 kids are outside of their school and they jump up with joy and the title of their school is then shown. A kid then takes a rock and throws it at a number on the hopscotch drawn on the ground with chalk. The video then begins by showing the steps to hopscotch and a boy begins drawing boxes and adding numbers up to the number 8. After, a marker is shown which happens to be a box and they throw it on the number 1 and then proceed to hop over the number one and come back and pick up the rock. The game continues and the line of boys throw the rock and take turns until the game is over.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_iDMzTPfELoc.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_hIJ6VTEKji8.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_8Kv7piYg9yc.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_lKKimizxQJM.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_msiX-xky6Ac.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 man is in a bathroom attempting to shave his legs. He takes off his pants, shakes the shaving cream can, sits on the toilet and puts his leg up. As other girls tease him about it he puts his leg up and lathers up his leg. He shaves his leg while the sink water runs then a girl touches the shaved area of his leg."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_NNOsdZr802w.mp4",
"activitynet/videos/v_rCSFBiXxbVs.mp4",
"activitynet/videos/v_2UfljrwzsLs.mp4",
"activitynet/videos/v_AO-0r8H2DOo.mp4",
"activitynet/videos/v_O2Y6rn4gFd4.mp4"
] | 1
|
<answer>1</answer>
|
[
1
] | true
| 1
| 7,532.804348
| 392
|
[
{
"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 man is in a bathroom attempting to shave his legs. He takes off his pants, shakes the shaving cream can, sits on the toilet and puts his leg up. As other girls tease him about it he puts his leg up and lathers up his leg. He shaves his leg while the sink water runs then a girl touches the shaved area of his leg.\"\n\nCandidates:\n[1] 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
}
},
{
"type": "text",
"text": "\n[2] 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[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_2UfljrwzsLs.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_AO-0r8H2DOo.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_O2Y6rn4gFd4.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 holds a decoration and decides where to place it on the Christmas tree. The woman hangs an ornament on the Christmas tree. An angel rests at the top of a Christmas tree. The woman picks up an ornament."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_8tlLBffNjf8.mp4",
"activitynet/videos/v_lRB6XvAm_FU.mp4",
"activitynet/videos/v_LYqq0dPB-U8.mp4",
"activitynet/videos/v_95UgspVYJSM.mp4",
"activitynet/videos/v_SIf4H2dqbpg.mp4"
] | 4
|
<answer>4</answer>
|
[
4
] | true
| 1
| 22,678.289467
| 395
|
[
{
"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 holds a decoration and decides where to place it on the Christmas tree. The woman hangs an ornament on the Christmas tree. An angel rests at the top of a Christmas tree. The woman picks up an ornament.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_8tlLBffNjf8.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_lRB6XvAm_FU.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_LYqq0dPB-U8.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_95UgspVYJSM.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_SIf4H2dqbpg.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 images flash then we see the title screen. We see a man walking a large group of dogs while talking to the camera. We see the man pouring into dish. The man rings a bell at a building. They walk down a city street together. We see paws go across the screen. We see the record screen and the man picking up poo. The man asks another man how many dogs he has. The man walks past a bunch of schoolgirls."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_whJ6ESGNoyY.mp4",
"activitynet/videos/v_6EWzgWd72Cs.mp4",
"activitynet/videos/v_vgdcVhRSa9E.mp4",
"activitynet/videos/v_jkaevzzYdP8.mp4",
"activitynet/videos/v_jEppv00aBBc.mp4"
] | 5
|
<answer>2</answer>
|
[
2
] | false
| 0
| 32,736.828274
| 394
|
[
{
"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 images flash then we see the title screen. We see a man walking a large group of dogs while talking to the camera. We see the man pouring into dish. The man rings a bell at a building. They walk down a city street together. We see paws go across the screen. We see the record screen and the man picking up poo. The man asks another man how many dogs he has. The man walks past a bunch of schoolgirls.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_whJ6ESGNoyY.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_6EWzgWd72Cs.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_vgdcVhRSa9E.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_jkaevzzYdP8.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_jEppv00aBBc.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 people in inflatable boats sailing in the river. One person gets pushed off of the boat, but the others keep going. when they get to shore, they see other people who they know. There is a boy cooking food over fire, one boy is peeling potatoes, and other things are happening."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_OabVylOVys4.mp4",
"activitynet/videos/v_yJSQmNSFlNI.mkv",
"activitynet/videos/v_FX4inHmWQtE.mp4",
"activitynet/videos/v_T3bTwmccIEQ.mp4",
"activitynet/videos/v_cT4EquMmRiw.mp4"
] | 1
|
<answer>2</answer>
|
[
2
] | false
| 0
| 34,194.116944
| 393
|
[
{
"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 people in inflatable boats sailing in the river. One person gets pushed off of the boat, but the others keep going. when they get to shore, they see other people who they know. There is a boy cooking food over fire, one boy is peeling potatoes, and other things are happening.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_OabVylOVys4.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_yJSQmNSFlNI.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[3] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_FX4inHmWQtE.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_T3bTwmccIEQ.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_cT4EquMmRiw.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A person is seen speaking to the camera while rubbing himself in sunscreen and showing off several different kinds of lotion. The man is shown speaking with his shirt on and off while rubbing sun screen into his skin and wiping his face with a towel."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_ChPzol03Hqs.mp4",
"activitynet/videos/v_mf6UsZuW9Nw.mp4",
"activitynet/videos/v_s0swzu1jIpc.mkv",
"activitynet/videos/v_ABQYqpWF1LA.mp4",
"activitynet/videos/v_6fWXqCWuU9Y.mp4"
] | 4
|
<answer>4</answer>
|
[
4
] | true
| 1
| 14,809.382535
| 396
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames.\n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A person is seen speaking to the camera while rubbing himself in sunscreen and showing off several different kinds of lotion. The man is shown speaking with his shirt on and off while rubbing sun screen into his skin and wiping his face with a towel.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_ChPzol03Hqs.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_mf6UsZuW9Nw.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_s0swzu1jIpc.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_ABQYqpWF1LA.mp4",
"nframes": 48,
"total_pixels": 50176
}
},
{
"type": "text",
"text": "\n[5] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_6fWXqCWuU9Y.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "There's a dog in a living that has a messy couch and an over crowded coffee table filled with books and TV remotes. A person walks by to lift something off the floor. The dog watches the person and leaves the room momentarily. Then the dog comes back and begins drinking from a coffee mug that is placed on the coffee table. The stands on its hind legs to reach the mug and drink the liquid in it. The dog jumps down and goes stands back up on its hind legs to drink some more from the mug. After the dog leaves the room, a person walks in."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_BSg989GP5ro.mp4",
"activitynet/videos/v_44FeihJUKvM.mp4",
"activitynet/videos/v_DAPX3S1Nmqg.mp4",
"activitynet/videos/v_1mYtNMDFyXQ.mkv",
"activitynet/videos/v_sF859t5osSg.mp4"
] | 5
|
<answer>5</answer>
|
[
5
] | true
| 1
| 16,905.809387
| 397
|
[
{
"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 dog in a living that has a messy couch and an over crowded coffee table filled with books and TV remotes. A person walks by to lift something off the floor. The dog watches the person and leaves the room momentarily. Then the dog comes back and begins drinking from a coffee mug that is placed on the coffee table. The stands on its hind legs to reach the mug and drink the liquid in it. The dog jumps down and goes stands back up on its hind legs to drink some more from the mug. After the dog leaves the room, a person walks in.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_BSg989GP5ro.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_44FeihJUKvM.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_DAPX3S1Nmqg.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_1mYtNMDFyXQ.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_sF859t5osSg.mp4",
"nframes": 48,
"total_pixels": 50176
}
}
]
}
] |
You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.
System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames.
Decision rules (concise):
- Judge relevance ONLY from what is visible in the frames.
- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.
- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.
ABSOLUTE OUTPUT CONTRACT:
<answer> [a] > [b] > [c] > [d] > [e] </answer>
Constraints:
- Do NOT use per-candidate XML tags.
- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.
- Always output a total order in <answer>.
Query: "A little boy is holding a tube of toothpaste. His mom helps him put it on his toothbrush. The boy then uses the brush to brush his teeth."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_ERGoTBC8NkA.mp4",
"activitynet/videos/v_1rdecGieY-M.mp4",
"activitynet/videos/v_3SLaaTD8t3Q.mp4",
"activitynet/videos/v_yu1XjQUctiM.mp4",
"activitynet/videos/v_rWDMssiL7hE.mp4"
] | 4
|
<answer>4</answer>
|
[
4
] | true
| 1
| 19,606.54897
| 399
|
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "You will receive a Query and N candidates. Each candidate is a video. Your task is to identify the most relevant video to the Query. Output only the index of the most relevant video among the candidates. Your answer should be strictly inside <answer></answer> tags.\n\nSystem Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames.\n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n\n<answer> [a] > [b] > [c] > [d] > [e] </answer>\n\n\nConstraints:\n- Do NOT use per-candidate XML tags.\n- a,b,c,d,e are indices of the candidates. you should output the number in the place of a,b,c,d,e.\n- Always output a total order in <answer>.\n\nQuery: \"A little boy is holding a tube of toothpaste. His mom helps him put it on his toothbrush. The boy then uses the brush to brush his teeth.\"\n\nCandidates:\n[1] 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[2] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_1rdecGieY-M.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_3SLaaTD8t3Q.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_yu1XjQUctiM.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_rWDMssiL7hE.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 male gymnast perform on the parallel bars. First the man stretches and sticks out his butt. He then mounts the parallel bars. The man misses a bar and falls. The man puts more chalk on his hands. He gets back on the bars and flips a few times before dismounting the bars."
Candidates:
[1] video:<video>
[2] video:<video>
[3] video:<video>
[4] video:<video>
[5] video:<video>
|
[
"activitynet/videos/v_7DJDUzdw_I4.mp4",
"activitynet/videos/v_QVdsLRKpCT0.mp4",
"activitynet/videos/v_mbGpp_nDwI4.mp4",
"activitynet/videos/v_dAcdSkaoK64.mp4",
"activitynet/videos/v_XFijgUPprk4.mp4"
] | 3
|
<answer>1</answer>
|
[
1
] | false
| 0
| 24,100.463438
| 398
|
[
{
"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 male gymnast perform on the parallel bars. First the man stretches and sticks out his butt. He then mounts the parallel bars. The man misses a bar and falls. The man puts more chalk on his hands. He gets back on the bars and flips a few times before dismounting the bars.\"\n\nCandidates:\n[1] video:",
"video_url": null
},
{
"type": "video_url",
"text": null,
"video_url": {
"url": "file:///hub_data3/seohyun/activitynet/videos/v_7DJDUzdw_I4.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_QVdsLRKpCT0.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_mbGpp_nDwI4.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_dAcdSkaoK64.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_XFijgUPprk4.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.