ReFoCUS: Reinforcement-guided Frame Optimization for Contextual Understanding
Paper • 2506.01274 • Published • 4
You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This dataset is provided for academic research and education only. Access is granted automatically once the form below is filled in.
Log in or Sign Up to review the conditions and access this dataset content.
ReFoCUS-393K metadata
Metadata of ReFoCUS-393K, the dataset used to train the ReFoCUS frame-selection policy (paper, code). The videos are available at interlive/ReFoCUS-393K.
Files
| File | Samples | Description |
|---|---|---|
pool.json |
393,216 | training pool |
train_stage1.json … train_stage4.json |
98,304 each | the pool split per curriculum stage |
reward_eval.jsonl |
961,690 | reward-model evaluations the pool was selected from |
From each source (LLaVA-Video-178K, CinePile, VISTA-400K), questions whose prediction
varies the most with the temporal window the model looks at, that is, those demanding
high temporal focus, were collected first, giving the 393,216 samples of pool.json;
train_stage<stage>.json splits them per stage.
For reward_eval.jsonl, InternVL3-2B was fed 32 uniformly sampled frames from each
temporal window of the video and the log-probability of the answer token was recorded.
See Sec. A.1, Fig. 8 of the paper.
Sample
{
"video": "cinepile/G3t30jAYDZA.mp4",
"question": "...",
"options": ["...", "...", "...", "...", "..."],
"answer": 2, // 1-based
"conversations": [{"from": "human", "value": "<image>\n..."}, {"from": "gpt", "value": "B. ..."}],
"clip_num": 1,
"scenes": [{"start_time": 0.0, "end_time": 166.5, "start_frame": 0, "end_frame": 666,
"start_fps": 4, "end_fps": 4, "duration": 166.5}]
}
Licenses
The labels derive from the source datasets and inherit their terms: LLaVA-Video-178K (Apache-2.0, academic use), CinePile (CC BY-NC-SA 4.0) and VISTA-400K (MIT).
- Downloads last month
- 20