README.md

metadata

license: mit
task_categories:
  - question-answering
tags:
  - chess
  - reasoning
  - strategy
  - tactics
  - llm
  - game-playing
language:
  - en
size_categories:
  - 100K<n<1M
pretty_name: MATE - Move on strAtegy and Tactics datasEt

MATE - Move on strAtegy and Tactics datasEt

Dataset Description

Dataset Summary

MATE (Move on strAtegy and Tactics datasEt) is a dataset of around 1 million chess positions designed to explore and enhance the reasoning capability of large language models in chess . Each position includes candidate moves annotated by chess experts (including world champion-level players) with detailed explanations of long-term strategy and short-term tactics .

The dataset is inspired by how expert chess players employ a dual approach combining long-term strategic planning with short-term tactical calculations along with language explanations .

Supported Tasks and Leaderboards

This dataset supports:

• Chess move selection and evaluation
• Strategic reasoning in game playing
• Tactical calculation and pattern recognition
• Language-guided decision making in complex reasoning tasks

Languages

The dataset is in English, with chess positions in FEN (Forsyth-Edwards Notation) format and moves in UCI (Universal Chess Interface) format.

Dataset Structure

The dataset contains four sub-datasets:

1. MATE-No-Explanation (40.8%): Chess positions with candidate moves but no explanations
2. MATE-Strategy (39.2%): Chess positions with candidate moves and strategic elaboration
3. MATE-Tactic (10%): Chess positions with candidate moves and tactical descriptions
4. MATE-Strategy&Tactic (10%): Chess positions with candidate moves and both strategy and tactic explanations

Data Fields

Each instance contains:

• position: Chess board position in FEN format
• moves: Candidate moves in UCI format (typically 2 moves per position)
• strategy (when applicable): Long-term strategic explanation categorized into:
  • Material count (6.5%)
  • Piece activity (65.2%)
  • Pawn structure (3.6%)
  • Space (8.4%)
  • King safety (16.3%)
• tactic (when applicable): Short-term tactical sequences with move descriptions, ranging from 1-6 moves

Data Splits

The dataset can be split into train/validation/test sets based on user requirements. Each sub-dataset maintains similar difficulty levels as validated through both human and automated assessment.

Curation Rationale

The dataset was created to study how strategic and tactical language explanations can guide language models to find better chess moves, addressing the limitation that language models show significant difficulties in planning and reasoning for complex tasks .

Initial Data Collection

The positions were collected from the open source chess server Lichess, including both chess games and chess puzzles . The collection follows these guidelines:

• Coverage of all game phases (opening, middlegame, endgame)
• Inclusion of different strategies and tactics
• Representation of various skill levels

Annotations

Annotation process

For each position, multiple reasonable moves were selected and annotated with language explanations by expert chess players, including world champion-level players .

Strategy Annotations: Experts formulated rules to determine optimal strategies for positions, with approximately 20 distinct linguistic expressions per strategic category .

Tactic Annotations: Move sequences were generated using the open source chess engine Stockfish, with factual descriptions of resulting positions .

Who are the annotators?

Chess experts including world champion-level players (notably, co-author Yifan Hou is mentioned as a four-time chess world champion).

Citation Information

@article{wang2024explore,
  title={Explore the reasoning capability of llms in the chess testbed},
  author={Wang, Shu and Ji, Lei and Wang, Renxi and Zhao, Wenxiao and Liu, Haokun and Hou, Yifan and Wu, Ying Nian},
  journal={arXiv preprint arXiv:2411.06655},
  year={2024}
}