---
license: other
tags:
- video
- reasoning
- maze
- chain-of-frames
---

# COF Maze raw training data

Raw Maze training data used by the COF Maze experiments.

## Contents

- 4,000 training samples: 1,000 each for 3x3, 4x4, 5x5, and 6x6 mazes
- 4,000 MP4 ground-truth videos
- 4,000 conditioning first-frame PNGs
- 4,000 topology/solution JSON files
- one `metadata.csv`
- videos: 81 frames, 480x480, 16 fps

The files are stored in `maze_train_raw/maze_train_raw_20260904.tar.zst`. The archive expands to a single
`maze_train/` directory with 12,001 files.

## Integrity

- archive size: `70962088` bytes
- SHA-256: `4df6980d5327b7700f0ca0f871cf56469f07b23e1d473a26d7fd2e21e0605d86`
- release manifest: `maze_train_raw_release.json`

## Download

```bash
wget https://huggingface.co/datasets/Boxue/cof-maze-training-data/resolve/main/maze_train_raw/maze_train_raw_20260904.tar.zst
echo '4df6980d5327b7700f0ca0f871cf56469f07b23e1d473a26d7fd2e21e0605d86  maze_train_raw_20260904.tar.zst' | sha256sum -c -
zstd -dc maze_train_raw_20260904.tar.zst | tar -xf -
```

The latent-preencoded LMDB is a derived artifact and is not duplicated in this
raw-data repository. Preserve upstream attribution and review the source
project's dependency licenses before redistribution or reuse.