Datasets:
init dataset card
Browse files
README.md
ADDED
|
@@ -0,0 +1,21 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: other
|
| 3 |
+
task_categories:
|
| 4 |
+
- text-to-image
|
| 5 |
+
- image-to-text
|
| 6 |
+
language:
|
| 7 |
+
- en
|
| 8 |
+
tags:
|
| 9 |
+
- webdataset
|
| 10 |
+
- video-frames
|
| 11 |
+
- captions
|
| 12 |
+
size_categories:
|
| 13 |
+
- 1M<n<10M
|
| 14 |
+
---
|
| 15 |
+
# Video-frame image/caption dataset
|
| 16 |
+
|
| 17 |
+
Frames sampled from user-provided videos (deterministic hash sampling), quality filtered and de-duplicated,
|
| 18 |
+
captioned with `Qwen/Qwen3-VL-8B-Instruct`. Stored as WebDataset shards (`data/train-*.tar`: `<id>.jpg` + `<id>.json`),
|
| 19 |
+
with per-shard Parquet indexes in `index/` and resume state in `control/`.
|
| 20 |
+
Load with `webdataset` (`wds.WebDataset("hf://datasets/<repo>/data/train-{000001..00000N}.tar")`) or stream the tar files.
|
| 21 |
+
Respect the licenses of the source videos (see `source_name` / `license` fields in each JSON).
|