metadata

dataset_info:
  features:
    - name: state
      dtype: string
    - name: action
      dtype: string
    - name: future
      dtype: string
  splits:
    - name: train
      num_bytes: 184646417
      num_examples: 659576
    - name: test
      num_bytes: 5755726
      num_examples: 62561
  download_size: 52539213
  dataset_size: 190402143
license: apache-2.0

Dataset Card for "chess10k"

Dataset Details

This is a chess dataset created in the paper "Implicit Search via Discrete Diffusion: A Study on Chess".

chess10k contains 10k games from Lichess and the actions are reannotated by Stockfish 16.

The datasets contains three fields:

state: the board state, which is represented in a FEN-like format.
action: the best action, which is suggested by Stockfish 16 at the above state.
future: the best future trajectory, which is suggested by Stockfish 16 followed by the above action.

Citation

@article{ye2025implicit,
  title={Implicit Search via Discrete Diffusion: A Study on Chess},
  author={Ye, Jiacheng and Wu, Zhenyu and Gao, Jiahui and Wu, Zhiyong and Jiang, Xin and Li, Zhenguo and Kong, Lingpeng},
  journal={arXiv preprint arXiv:2502.19805},
  year={2025}
}