theoden8/nnue-chess-dataset

---
license: cc0-1.0
task_categories:
- tabular-regression
pretty_name: NNUE Chess Dataset (dummy_chess)
---

# NNUE Chess Dataset

This dataset contains chess positions with evaluations for training NNUE (Efficiently Updatable Neural Network) chess engines.

## Dataset Composition

The dataset consists of three distinct data sources, each providing different types of positions and evaluation contexts:

### 1. Evaluations
- **Source**: [Lichess Open Database](https://database.lichess.org/)
- **Method**: Positions extracted from Lichess games with their corresponding engine evaluations
- **Evaluation**: First principal variation (PV) line score from the original analysis
- **Purpose**: Provides diverse positions from real games across various skill levels and opening types

### 2. Puzzles
- **Source**: [Lichess Open Database](https://database.lichess.org/)
- **Method**: Extracts two critical positions from each puzzle:
- Pre-mistake position (before the critical error)
- Post-mistake position (after the critical error)
- **Evaluation**: Re-analyzed using Stockfish 16 (depth 12) with tablebase support (5-man)
- **Purpose**: Focuses on tactically rich positions where accurate evaluation is critical

### 3. Endgames
- **Source**: Tablebase positions
- **Method**: Sampled positions from endgame tablebases
- **Evaluation**: Scored using Stockfish 16 (depth 12) with tablebase support (5-man)
- **Purpose**: Ensures strong endgame understanding with theoretically perfect tablebase knowledge

## Technical Specifications

### Engine Configuration
- **Stockfish Version**: 16
- **Search Depth**: 12 plies
- **Tablebase**: [Syzygy 5-man endgame tablebases](https://github.com/syzygy1/tb)
- **Tablebase Depth**: 5-man positions

### Data Format
- **Position Format**: Positions are stored in the `fen` column using a compressed FEN notation
- **Decompression**: Use the `decompress_fen` function from [dummy_chess](https://github.com/theoden8/dummy_chess) to convert to standard FEN
- **Evaluation Format**: All positions include centipawn (cp) evaluations from White's perspective, or mate-in-N scores where applicable

## Tools & Dependencies

This dataset was created using the following tools:

- **[dummy_chess](https://github.com/theoden8/dummy_chess)**: Used for preprocessing, PGN parsing, position extraction, and FEN compression/decompression
- **[python-chess](https://github.com/niklasf/python-chess)**: Chess library for position handling and move generation
- **[Stockfish 16](https://stockfishchess.org/)**: Engine for position evaluation
- **[Syzygy Tablebases](https://github.com/syzygy1/tb)**: Endgame tablebases for perfect endgame knowledge

## Data Sources

- [Lichess Open Database](https://database.lichess.org/) - Games and evaluations
- Lichess Puzzle Database - Tactical positions
- Syzygy Tablebases - Endgame positions

## Use Cases

This dataset is suitable for:
- Training NNUE networks for chess engines (specifically [dummy_chess](https://github.com/theoden8/dummy_chess))
- Supervised learning for position evaluation
- Testing evaluation function accuracy
- Research in chess AI and position assessment

## License

[Add appropriate license information]

## Citation

If you use this dataset, please cite the relevant sources:
- Lichess Open Database: https://database.lichess.org/
- dummy_chess: https://github.com/theoden8/dummy_chess

## Acknowledgments

This dataset was created for the [dummy_chess](https://github.com/theoden8/dummy_chess) project.