Ramora0/chess-av-mates

ChessBench Action-Values + Mate-in-N

This dataset is derived from the action-value data released with "Grandmaster-Level Chess Without Search" (DeepMind). It provides:

1. A reorganization of the original action-value format into a per-position structure: each FEN maps to a list of all legal moves with per-move win probabilities (as provided in the upstream release).
2. A mate-in-N augmentation: for moves with win probability >= 99.99% or <= 0.01%, Stockfish mate search adds a mate depth field per move when a forced mate is detected.

Use Cases

• Training a policy model with full stockfish signal, rather than just best move as Deepmind does
• Training models to predict mate imminence signals beyond saturated win rates

Source Dataset

• Upstream: DeepMind "searchless chess" / ChessBench action-value release
• Paper: "Grandmaster-Level Chess Without Search" (Ruoss et al., 2024)
• Upstream license: Some portions are CC0 (Lichess), remainder is CC BY 4.0

Schema

Each row contains:

Field	Type	Description
fen	str	Chess position in FEN notation
moves	List[str]	All legal moves in UCI format (e.g., ["e2e4", "d2d4", ...])
p_win	List[float]	Win probability for side-to-move per move, in [0.0, 1.0] (unchanged from source)
mate	List[int]	Mate depth per move (new field, see below)

All three list fields (moves, p_win, mate) share the same length for each row and are correlated.

Mate Field Definition

The mate field encodes forced mate depth in full moves (not plies), from the perspective of the side to move:

Value	Meaning
mate > 0	Playing this move leads to mating the opponent in N moves
mate < 0	Playing this move leads to getting mated in N moves
mate = 0	No forced mate detected, or move was not analyzed

Only moves with stockfish win probability of 0% or 100% are analyzed, since these represent when stockfish found a mate in its search. All other moves have mate = 0 by default.

Analysis Thresholds

• Moves with p_win >= 0.9999 are analyzed as potential winning mates
• Moves with p_win <= 0.0001 are analyzed as potential losing mates
• All other moves are not analyzed (mate = 0)

How Mate-in-N Was Computed

Mate depth labels were generated using Stockfish with depth 16; we found this found forced mates for ~98.5% of moves with winrates of 0% or 100%. It also means the maximum mate depth is 8 moves; when move winrate is 0% or 100% and mate is 0, it likely means a deeper stockfish found a forced mate in 8+ moves.

Analysis Procedure

For each move meeting the win probability threshold:

1. The move is applied to the board position
2. If the resulting position is immediate checkmate, mate = 1 (or -1 if a losing move)
3. Otherwise, Stockfish analyzes the resulting position to depth 16
4. If Stockfish reports a mate score, the mate depth (in full moves) is recorded, + 1 to account for our original move
5. If no mate score is found within the depth limit, mate = 0

Note: mate = 0 does not imply no forced mate exists -- only that none was found within the search depth limit.

Limitations / Known Issues

• Mate-in-N labels depend on Stockfish search depth; deeper searches may find mates that depth 16 misses
• Upstream win probabilities near 0% or 100% may reflect Stockfish evaluation saturation rather than true forced mates, which is why not all analyzed moves yield confirmed mates
• Legal move generation and UCI formatting must match python-chess rules for underpromotions, castling rights, and en passant

License

This dataset is licensed under CC BY 4.0.

Attribution & Changes

• Upstream attribution: DeepMind "searchless chess" action-value release
• Changes: Reorganization from per-move records into per-position legal-move lists, plus mate-in-N augmentation via Stockfish analysis

Citation

If you use this dataset, please cite the upstream DeepMind paper:

@article{ruoss2024grandmaster,
  title={Grandmaster-Level Chess Without Search},
  author={Ruoss, Anian and Del{\'e}tang, Gr{\'e}goire and Medapati, Sourabh and Grau-Moya, Jordi and Wenliang, Li Ke and Catt, Elliot and Reid, John and Genewein, Tim},
  journal={arXiv preprint arXiv:2402.04494},
  year={2024}
}