hannahyh/vlm-distill-gigahands-brics001-cam0

---
pretty_name: VLM Distill GigaHands BRICS001 Cam0
task_categories:
- robotics
- video-classification
tags:
- robotics
- vla
- distillation
- gigahands
- hand-keypoints
---

# VLM Distill GigaHands BRICS001 Cam0

Training and testing data used for VLA-HAND VLM distillation experiments on the GigaHands real dataset, restricted to BRICS001 camera 0 keypoint-linked clips.

The original data contains more than 10,000 files in one annotation directory, which exceeds the Hugging Face Hub per-directory file limit. The dataset is therefore stored as archives that preserve the original directory layout when extracted.

## Files

- `archives/annotation_gigahands_real_train.tar.zst`: `Annotation/gigahands_real_train/`
- `archives/annotation_gigahands_real_test.tar.zst`: `Annotation/gigahands_real_test/`
- `archives/video_gigahands_root.tar.zst.part-*`: split archive for `Video/GigaHands_root/`
- `subset_manifest.json`: source subset manifest.
- `conversion_report.json`: conversion summary.
- `gigahands_real_train_keypoints_weighted_statistics.json`: train split statistics.
- `gigahands_real_test_keypoints_weighted_statistics.json`: test split statistics.
- `needed_videos_full_keypoints_brics001_cam0_unique.txt`: source video list used to stage this upload.

## Restore

```bash
mkdir -p restored
tar --use-compress-program=unzstd -xf archives/annotation_gigahands_real_train.tar.zst -C restored
tar --use-compress-program=unzstd -xf archives/annotation_gigahands_real_test.tar.zst -C restored
cat archives/video_gigahands_root.tar.zst.part-* > /tmp/video_gigahands_root.tar.zst
tar --use-compress-program=unzstd -xf /tmp/video_gigahands_root.tar.zst -C restored
```

After extraction, use `restored` as the dataset root. The corresponding local VLA-HAND config used `data_mix` values `gigahands_real_train` for training and `gigahands_real_test` for evaluation/testing.