---
license: other
task_categories:
- text-to-image
- image-to-text
language:
- en
tags:
- webdataset
- video-frames
- captions
size_categories:
- 1M<n<10M
---
# Video-frame image/caption dataset

Frames sampled from user-provided videos (deterministic hash sampling), quality filtered and de-duplicated,
captioned with `Qwen/Qwen3-VL-8B-Instruct`. Stored as WebDataset shards (`data/train-*.tar`: `<id>.jpg` + `<id>.json`),
with per-shard Parquet indexes in `index/` and resume state in `control/`.
Load with `webdataset` (`wds.WebDataset("hf://datasets/<repo>/data/train-{000001..00000N}.tar")`) or stream the tar files.
Respect the licenses of the source videos (see `source_name` / `license` fields in each JSON).