VideoSearcher RL Training Data
RL training/validation data for VideoSearcher (video deep-research agent), as used by
train_video_test_GSPO_HDPO_image.sh. All paths inside the annotation files are
relative to this directory.
Contents
| File / Dir | Description |
|---|---|
train_video_deepresearch_qcap1_plus_0505.json |
Train dataset config (points to the train jsonl) |
test_video_deepresearch.json |
Val dataset config (points to the val jsonl) |
all_video_deepresearch_qcap1_plus_video_dr_0505_rl.jsonl |
Train annotations, 3285 samples |
val_video_deepresearch_data.le20min.jsonl |
Val annotations, 348 samples |
videos/<id>/all_frames/frame_*.jpg |
1fps frames for fvqa videos (2324 videos) |
frames_dr0505/<id>/frame_*.jpg |
1fps frames for video_dr_0505 videos (1124 videos) |
Each jsonl line has an image field listing frame paths (relative), e.g.
videos/fvqa_1002__yUuVs1AMjHk/all_frames/frame_000000.jpg or
frames_dr0505/vdr_015_.../frame_000000.jpg.
Frame index (1fps) is watermarked in the top-left corner of every frame.
Usage
Unzip all parts into one directory. To train, point the dataset configs' annotation
field (or your data root) at this directory, or prepend this directory's absolute path
to the relative paths in the jsonl image fields.