VSeek
Collection
Artifacts for the ECCV 2026 Oral Paper: Incentivizing Vision Language Models to Search for Long Video Question Answering. • 4 items • Updated
Datasets:
data_source string | prompt list | ability string | reward_model dict | extra_info dict |
|---|---|---|---|---|
lvb | [{"content":"\nYou are a video analysis agent that would aim to answer the user's question over mult(...TRUNCATED) | video_reasoning | {
"ground_truth": "1",
"style": "exact_match"
} | {"candidates":["promisc has come to an end, in and run away countless times, i was just scared, i st(...TRUNCATED) |
lvb | [{"content":"\nYou are a video analysis agent that would aim to answer the user's question over mult(...TRUNCATED) | video_reasoning | {
"ground_truth": "1",
"style": "exact_match"
} | {"candidates":["Music","Europe Russia China, India and Australia most of these lines actually don't (...TRUNCATED) |
lvb | [{"content":"\nYou are a video analysis agent that would aim to answer the user's question over mult(...TRUNCATED) | video_reasoning | {
"ground_truth": "3",
"style": "exact_match"
} | {"candidates":["Europe Russia China, India and Australia most of these lines actually don't exist. S(...TRUNCATED) |
lvb | [{"content":"\nYou are a video analysis agent that would aim to answer the user's question over mult(...TRUNCATED) | video_reasoning | {
"ground_truth": "2",
"style": "exact_match"
} | {"candidates":["cuts and bruises and they did not surrender","cuts and bruises and years that follow(...TRUNCATED) |
lvb | [{"content":"\nYou are a video analysis agent that would aim to answer the user's question over mult(...TRUNCATED) | video_reasoning | {
"ground_truth": "1",
"style": "exact_match"
} | {"candidates":["beath bars are mike chocolate covered","cooling for about maybe 10 minutes so and le(...TRUNCATED) |
lvb | [{"content":"\nYou are a video analysis agent that would aim to answer the user's question over mult(...TRUNCATED) | video_reasoning | {
"ground_truth": "3",
"style": "exact_match"
} | {"candidates":["history","develop","warfare","people","science"],"correct_choice":"3","index":5,"met(...TRUNCATED) |
lvb | [{"content":"\nYou are a video analysis agent that would aim to answer the user's question over mult(...TRUNCATED) | video_reasoning | {
"ground_truth": "0",
"style": "exact_match"
} | {"candidates":["that afternoon stella and her clay","let's find a cozy spot and let's get","let's be(...TRUNCATED) |
lvb | [{"content":"\nYou are a video analysis agent that would aim to answer the user's question over mult(...TRUNCATED) | video_reasoning | {
"ground_truth": "0",
"style": "exact_match"
} | {"candidates":["\"Music\"","\"built here since 16th century\"","\"places\"","\"this title\"","\"and (...TRUNCATED) |
lvb | [{"content":"\nYou are a video analysis agent that would aim to answer the user's question over mult(...TRUNCATED) | video_reasoning | {
"ground_truth": "4",
"style": "exact_match"
} | {"candidates":["\"sometimes we feel that our problems are\"","\"unbearable\"","\"with their own prob(...TRUNCATED) |
lvb | [{"content":"\nYou are a video analysis agent that would aim to answer the user's question over mult(...TRUNCATED) | video_reasoning | {
"ground_truth": "1",
"style": "exact_match"
} | {"candidates":["all right that’s the fun part of the day","is just bam it’s kind of slaps","Is t(...TRUNCATED) |
End of preview. Expand in Data Studio
VSeek data
Tag-summary question files used with the VSeek-EM and VSeek-VETL checkpoints.
train.parquet has 4,968 examples from LongVideoBench (lvb), MLVU (mlvu), and Video-MME (videomme).
test.parquet has 2,153 examples from those three datasets plus CG-Bench (cgbench) and LVBench (lvbench).
| split | lvb | mlvu | videomme | cgbench | lvbench |
|---|---|---|---|---|---|
| train | 1,069 | 1,739 | 2,160 | ||
| test | 268 | 435 | 540 | 600 | 310 |
- Downloads last month
- 20