metadata
configs:
- config_name: default
data_files:
- split: train
path: train/*.pgn
- split: test
path: test/*.pgn
license: mit
tags:
- code
size_categories:
- 100M<n<1B
pretty_name: ChessBot-Dataset
dataset_info:
features:
- name: state
dtype:
array2_d:
shape:
- 8
- 8
dtype: int8
- name: action
dtype: int16
- name: result
dtype: int8
splits:
- name: train
num_bytes: 65545518162
num_examples: 636364254
- name: test
num_bytes: 7298385021
num_examples: 70858107
download_size: 7621976329
dataset_size: 72843903183
This repository contains a gigantic curated chess dataset meant for machine learning. Currently the dataset contains approximately 700 million positions in PGN format, split across 1000 files. Huge credits to the following main sources:
- Lumbra's Database (filtered 2600+)
- Lichess Puzzle Database
- Computer Chess: TCEC Database, CCRL
Some stats (subject to change):
- Games per file: 9808/9809
- Positions per file: ~700,000
- Total train positions: 636364254
- Total test positions: 70858107
The supporting dataset code and training platform, known as ChessBot-Battleground, is found at: https://github.com/KeithG33/ChessBot-Battleground/
Note: The dataset requires trust_remote_code=True to parse the pgn data. See the code in ChessBot-Dataset.py if you'd like to verify first