---
license: cc0-1.0
language: en
pretty_name: ChessModel-XPU formal teacher dataset
tags:
  - chess
  - alphazero
  - stockfish
  - knowledge-distillation
  - game
  - reinforcement-learning
size_categories:
  - 1M<n<10M
---

# ChessModel-XPU formal teacher dataset

Stockfish-18-labeled chess positions used to train the
[`jinshuoli/chessmodel`](https://huggingface.co/jinshuoli/chessmodel) checkpoint,
part of the [ChessModel-XPU](https://github.com/JinShuo-Li/ChessModel) project.

It is a **custom-format** dataset of bit-packed board tensors plus sparse
Stockfish-derived policy / WDL targets. It is **not** loadable with
`datasets.load_dataset(...)`; use the project's `TeacherDataset` loader (see
below). Game-level train/validation split — adjacent positions from the same game
are never spread across splits.

## Source and provenance

| | |
|---|---|
| Source games | [Lichess Elite database, December 2023](https://database.lichess.org/) (`lichess_elite_2023-12`), licensed **CC0 1.0** |
| Source SHA-256 | `a6a5a8253cf357d31b7b5c1895a63dfbf64cc93b4504398f748cb69a31c0eff0` |
| Teacher | Stockfish 18 — MultiPV 8, 10000 nodes/position, WDL enabled, temperature 0.15 |
| Split | Deterministic game-level 90 / 10 split, seed 7, minimum 16 plies per game |
| Prep script | [`scripts/prepare_formal_pgn.py`](https://github.com/JinShuo-Li/ChessModel/blob/main/scripts/prepare_formal_pgn.py) |

Preparation removed parse failures, non-standard start positions, games without a
decisive/draw result, short games (< 16 plies), and duplicate move sequences, then
made the deterministic game-level split.

## Contents

The dataset mirrors the repository layout, so downloading it into a clone of the
project makes the workflow commands resolve unchanged:

| Path | Description |
|---|---|
| `data/formal_1m_train/shard-00000…00244.npz` | 245 training shards |
| `data/formal_50k_validation/shard-*.npz` | 13 validation shards |
| `datasets/formal_train.pgn` | Source training PGN (game-level split) |
| `datasets/formal_validation.pgn` | Source validation PGN (game-level split) |
| `datasets/formal_pgn_metadata.json` | Provenance metadata (source hash, counts, split params) |

Total ≈ 370 MB.

## Statistics

From `datasets/formal_pgn_metadata.json`:

| | |
|---|---|
| Parsed games | 315,135 |
| Accepted games | 312,603 |
| Train / validation games | 281,003 / 31,600 |
| Train / validation positions | 2,971,862 / 334,836 |
| Validation percent | 10 |
| Seed | 7 |

## Shard format

Each `.npz` shard is a project-defined record containing:

- Bit-packed board planes (`112 × 8 × 8`, canonically oriented to the side to move:
  up to 8 history frames, castling rights, en-passant, side to move, halfmove and
  fullmove clocks).
- Sparse policy target over the AlphaZero `8×8×73 = 4672` move encoding.
- Win/Draw/Loss target derived from Stockfish WDL.
- A shard format version, per-shard SHA-256 integrity check, Stockfish/node
  metadata, and the FEN (for legal-move masking and audit).

`TeacherDataset` loads each shard's bit-packed boards and sparse targets once and
unpacks lazily, so the full million-position dataset stays out of resident memory.

## How to load

Clone the repo and download the dataset into it, then use the project loader:

```bash
git clone https://github.com/JinShuo-Li/ChessModel.git
cd ChessModel
hf download jinshuoli/chessmodel-data --repo-type dataset --local-dir .
```

```python
from torch.utils.data import DataLoader
from chess_ai.data.dataset import TeacherDataset

train = TeacherDataset("data/formal_1m_train")          # 245 shards
val   = TeacherDataset("data/formal_50k_validation")     # 13 shards
loader = DataLoader(train, batch_size=512, shuffle=True)
```

Verify integrity with the project's verifier:

```bash
python scripts/verify_teacher_dataset.py --dataset data/formal_1m_train
python scripts/verify_teacher_dataset.py --dataset data/formal_50k_validation
```

## Intended use and limitations

- **Intended:** training / evaluating compact neural chess models via Stockfish
  distillation, and reproducing the project's training pipeline.
- **Custom format:** not consumable by the HF dataset viewer or
  `datasets.load_dataset()`; requires the project's loader.
- **Labels are Stockfish outputs:** policy/WDL targets reflect Stockfish 18 search
  at the configured node budget, not human game outcomes (game results are used
  only for splitting/filtering).

## License and attribution

- Source game data: © Lichess, **CC0 1.0** (public domain).
- The loading code and preparation scripts are MIT-licensed in the
  [source repository](https://github.com/JinShuo-Li/ChessModel).
- Stockfish is used solely as a teacher/labeling tool and is **not** distributed
  here.

## Related

- Trained checkpoint: [`jinshuoli/chessmodel`](https://huggingface.co/jinshuoli/chessmodel)
- Source code & documentation: [github.com/JinShuo-Li/ChessModel](https://github.com/JinShuo-Li/ChessModel)