chesslm-maia3-5m-stockfish-200k

ChessLM Maia3-5M Stockfish 200k

ChessLM is a 5.23M-parameter Maia3 chess-policy checkpoint fine-tuned on 200,000 public Stockfish-labelled Lichess positions. It starts from UofTCSSLab/Maia3-5M, preserves Maia3's chess-native square-token representation and legal-move policy, and adapts that policy toward deeper engine-labelled choices.

This is a transparent learning/research release. It is not a replacement for Stockfish, does not perform tree search, and has no calibrated Elo estimate.

Artifact identity

Training data

Training used 200,000 unique legal positions streamed from the public Lichess evaluation database. Each target was the first legal move of a qualified Stockfish principal variation at depth 20 or greater. Exact-FEN overlap between train, validation, and final holdout was zero.

This release contains no Lichess position dump. It contains only the model, compact experiment metadata, and evaluation reports.

Final holdout

The frozen checkpoint and untouched parent were evaluated once on the same 1,000-position holdout with Stockfish 18 at 20,000 nodes per evaluated move.

Mate accuracy regressed and is intentionally reported. On the shared 200k development validation set, the earlier 50k checkpoint also had better average CPL, blunder rate, and mate accuracy, while 200k had better top-1/top-3.

Corrected game evidence

Every opening was played twice with colors reversed, games were capped at 200 plies, and Stockfish used a fixed 1,000-node budget:

• versus ChessLM 50k: 7 wins, 7 draws, 6 losses — 52.5% score;
• versus Stockfish: 0 wins, 8 draws, 12 losses — 20.0% score;
• ChessLM 50k versus the same Stockfish budget: 0 wins, 4 draws, 16 losses — 10.0% score.

The 20% result came entirely from draws. It is not an Elo measurement and does not mean ChessLM beat Stockfish.

Installation and loading

Use Python 3.11 and install the pinned Maia3 implementation:

python3.11 -m venv .venv
.venv/bin/python -m pip install \
  "maia3 @ git+https://github.com/CSSLab/maia3.git@1e13597c42d4858b7cfd7cfdae01e297263364b2" \
  torch==2.13.0 python-chess==1.999 numpy==2.4.6 huggingface-hub==1.26.0

Download this repository, then load the explicit checkpoint with the included maia3_policy.py helper:

from pathlib import Path

from maia3_policy import load_policy_model

model, model_args = load_policy_model(
    Path("chesslm-maia3-5m-stockfish-200k.pt"),
    device="cpu",  # use "mps" on a compatible Apple Silicon Mac
)

The project repository's API demonstrates FEN validation, legal masking, and top-three move prediction. The .pt file is a PyTorch state dictionary; load model files only from sources you trust.

Intended use

• research and teaching about chess-policy adaptation;
• legal-move ranking from valid chess positions;
• controlled comparison with the untouched Maia3 parent; and
• local chess interfaces where limitations are visible.

Limitations

• Evaluation uses one target move even when several moves may be similarly strong.
• Each FEN is repeated across eight history slots. The model does not receive the real preceding game history.
• Self and opponent Elo inputs are fixed at 1500.
• The model selects from its policy in one forward pass and performs no search.
• Training emphasizes Stockfish agreement, not calibrated playing strength.
• The 200k checkpoint regressed on final mate accuracy and on several quality metrics versus the 50k checkpoint on shared validation positions.
• Results are specific to the recorded datasets, seeds, and Stockfish budgets.

Attribution and license

ChessLM is derived from Maia3-5M and uses the Maia3 implementation from the University of Toronto CSSLab. Maia3 code is released under GNU AGPL v3. This release includes the same license and preserves the upstream attribution.

Please cite the original Chessformer work:

@inproceedings{monroe2026chessformer,
  title={Chessformer: A Unified Architecture for Chess Modeling},
  author={Daniel Monroe and George Eilender and Philip Chalmers and Zhenwei Tang and Ashton Anderson},
  booktitle={The Fourteenth International Conference on Learning Representations},
  year={2026},
  url={https://openreview.net/forum?id=2ltBRzEHyd}
}

ChessLM experiment and adaptation work: Muhammad Suleman (msulemans), 2026.