chess-sf-labels

Stockfish-labelled chess positions with real game histories, for training and evaluating neural chess engines. Each record is one position from a rated Lichess game between strong players, annotated with Stockfish's best move, its evaluation in centipawns, its win/draw/loss model output, and its top-5 move ranking.

This is the training data behind the [LFM2.5-230M-Chess family of fine-tunes] (https://huggingface.co/mlabonne/LFM2.5-230M-Chess) — a 1M-position run of the same recipe took that model from ≈1751 to ≈1956 Elo (depth-3 search vs Stockfish 1800). Larger configs (games-10m, drills) are planned as the campaign behind them completes.

Config: games-1m

Split	Positions	Notes
train	998,457	1,543 additional positions timed out at depth 20 and were dropped
validation	2,000	held out before labelling; disjoint from train by encoded position

Source positions

Lichess/standard-chess-games (CC0), July–September 2025. Filters: rated Rapid or Classical, both players ≥ 2000 Elo, normal termination, at least 30 moves. Two positions per game sampled uniformly between ply 12 and two plies before the end (median ply 48 — mid-game heavy; deeper endgame coverage is a known gap, addressed in the planned larger configs). Deduplicated by the full encoded position+history; positions used in published benchmarks were excluded.

Labels

Generated with the official Stockfish 19 binary (linux universal), single thread, 64 MB hash, fresh engine per position (ucinewgame between positions):

• best, cp — Stockfish's move and score at depth 20 (centipawns for the side to move; forced mates clipped to ±10000)
• top — the MultiPV-5 ranking: up to five moves with their scores
• score — Stockfish's WDL expected score for the side to move, (wins + ½·draws)/1000 from UCI_ShowWDL
• bin — score mapped to the 64 win-probability bins used by the LFM2.5-230M-Chess prompt format (<v:k> covers [k/64, (k+1)/64))

Record schema

{"moves": ["e2e4", ...], "best": "g1f3", "cp": 34, "score": 0.53, "bin": 33,
 "depth": 20, "top": [{"uci": "g1f3", "cp": 34}, {"uci": "d2d4", "cp": 12}]}

moves is the UCI move list from the standard starting position; replaying it reaches the labelled position with its full history (the last 8 plies are part of the model prompt). All other fields are from the side to move's perspective.

Using it

The labels are model-agnostic — any engine that eats (position, best move, value, ranking) pairs can train on them. For the LFM2.5-230M-Chess prompt format (80-token board encoding, one token per move, <v:k> value bins), see that model's card; the recipe that consumed this exact config is described in the fine-tune's card.

Notes and limitations

• Depth-20 Stockfish labels are strong but not ground truth; ~0.15% of positions exceeded the 120 s labelling budget and are simply absent.
• cp/score distributions are game-realistic (median +5 cp; 10–90% range −622…+677): the set skews toward decisive, human-blunder-containing positions, which is exactly what real-play training wants but not a uniform position distribution.
• Positions are CC0 (Lichess); labels were machine-generated and follow the same publishing practice as Lichess's own [chess-position-evaluations] (https://huggingface.co/datasets/Lichess/chess-position-evaluations) and DeepMind's ChessBench.

Credits

• Lichess — the games, CC0
• Stockfish — the labelling engine (GPLv3 software, used as a tool; not distributed here)
• mlabonne/LFM2.5-230M-Chess — the model family this data was built for, and the prompt/token format reference