chsm8-pt

CHSM8

CHSM8 is a research chess move-prediction model. Given a game prefix, it predicts the next chess move while applying a legal-move mask at inference. It is a project-specific PyTorch model, not a standard Transformers model.

Source code, architecture, training scripts, and data pipeline: github.com/nov3o/llm_chess.

Architecture

• Llama-style causal chess decoder: 12 layers, width 960, 15 attention heads, SwiGLU MLP (width 3,520), RoPE, and Q/K normalization.
• Factorized chess output heads predict move kind, source square, destination square, piece, and promotion rather than a flat move vocabulary.
• About 165M trainable parameters. The architecture retains a frozen dummy cross-attention scaffold for compatibility with later text-conditioned work.
• Trained in bf16 with FlashAttention, fused AdamW, gradient clipping 1.0, and weight decay 0.1.

Training data and schedule

Training data: LegumMagister/chsm8-pre-training (1.07B filtered Lichess games; the dataset's scripts/ rebuild the exact packed format). Training uses packed Lichess standard-chess game records, represented as move sequences. The final branch reads the 3p-s packed corpus without replacement in 512-position sequences; shards 77–127 were globally reshuffled during training to remove a temporal data-order artefact.

The planned one-pass budget is 78.32B chess positions. Optimisation uses a position-indexed cosine learning-rate schedule: 5% warmup, peak LR 5e-4, and minimum LR equal to 10% of the peak. A 1% held-out split is reserved for validation. The model's legal-move penalty weight is 0.5.

Checkpoint

A weights-only PyTorch dictionary containing model_state_dict, step, total_training_time, total_tokens_processed, and cfg. Load it with FactoredChessDecoder / DecoderConfig from this repository's src/model.py.

Evaluation

Estimated strength: 1837 Elo (95% CI 1817–1856) on Stockfish's UCI_Elo scale.

* 1,000 games played; one reached the 240-ply cap without a result and is excluded.

Protocol: colour-balanced games, greedy legal-move decoding (the model always plays its highest-scoring legal move; no search, no sampling), against Stockfish 17.1 with UCI_LimitStrength / UCI_Elo=2000, 0.05 s per move, 1 thread, 240-ply cap. Elo = 2000 + 400·log10(S / (1 − S)), with S the score (win = 1, draw = ½); the interval comes from the per-game standard error. UCI_Elo is calibrated against engine rating lists, not Lichess or FIDE ratings, so treat the number as a reproducible project metric.

Checkpoint choice. We originally expected the 85%-of-training checkpoint to be the better model, because it precedes a spike in training loss late in the run, and it did lead after a 200-game evaluation. Over 1,000 games, however, this final checkpoint came out ahead, so it is the one published.

Comparison with other chess models

Head-to-head, 200 games per opponent: 100 book openings, each played with both colours; both sides play their single most likely legal move (no search, no sampling); games past 178 plies count as draws (3 of 1,000).

All compared models, like chsm8-pt, are trained only by supervised next-move / next-token prediction, with no reinforcement learning, reward models or preference optimization.

Limitations

CHSM8 is a research checkpoint with no search, opening book, endgame tablebase, or standard Transformers interface. It has been evaluated only under the project protocol above; use it for research and reproduce results from the linked codebase.