# Benchmark Hugging Face Release

This directory contains the current benchmark release split into question JSON files and per-community data archives.

- `datasets/codabench.json`: full benchmark questions.
- `datasets/codabench-hard.json`: hard subset questions.
- `archives/community_*.tar.zst`: compressed data archives for Hugging Face Git LFS upload.
- `archives_manifest.json`: archive size, SHA256, and extraction target metadata.
- `release_manifest.json`: source-to-release community mapping and copy statistics.
- `scripts/extract_archives.py`: verifies archive SHA256 values and extracts data.

Each question row has `release_community` and `data_path`. After extracting an archive, its contents land at that `data_path`, for example:

```bash
python scripts/extract_archives.py --community community_0
```

For all archives:

```bash
python scripts/extract_archives.py
```

The data copy dereferences the symlinks produced by `stage1_setup`, so the archives contain actual file contents rather than symlink placeholders.

Recommended Hugging Face layout:

```text
datasets/
archives/
archives_manifest.json
release_manifest.json
README.md
scripts/
```