ChessModel-XPU — formal_1m checkpoint

A compact policy / WDL residual network for chess, trained by distilling Stockfish 18 MultiPV / WDL analysis. This is the formal_1m checkpoint from the ChessModel-XPU project — the main_xpu / main_cuda preset (12 residual blocks × 192 channels, 8,932,076 parameters, squeeze-and-excitation with hidden 32).

It is intended to be used together with chess rules and batched PUCT search as a neural chess engine. The network alone does not play chess; the playable engine runs MCTS-style PUCT over it. See the source repository for the search code.

Research artifact. The project optimizes for strength per unit compute, not absolute strength, and makes no Elo claim for this checkpoint. Measured strength is reported through the project's paired Stockfish evaluation matches, not asserted here.

Model details

Architecture	BatchNorm residual tower, spatial 73-plane policy head, 3-logit Win/Draw/Loss head (no moves-left head in this checkpoint)
Preset	main_xpu (main_cuda on the Linux/CUDA branch) — 12 blocks × 192 channels, SE hidden 32
Parameters	8,932,076
Input	112 × 8 × 8 planes, canonically oriented to the side to move (8 history frames, castling rights, en-passant, side to move, halfmove/fullmove clocks)
Policy output	AlphaZero 8×8×73 = 4672 move encoding (56 queen rays, 8 knight moves, 9 underpromotions); illegal logits are masked
Value output	3 logits → Win / Draw / Loss
License	MIT (see LICENSE)
Developer	JinShuo-Li

Training

Distilled from Stockfish 18 teacher labels; trained on the companion dataset jinshuoli/chessmodel-data.

Key settings from configs/formal_1m.yaml:

Batch size	512
Epochs	20
Learning rate	1e-3, cosine schedule, 2000 warmup steps
Weight decay	1e-4
Loss	policy 1.0 + value 1.0 (Win/Draw/Loss)
Precision	BF16 autocast
Teacher	Stockfish 18, MultiPV 8, 10000 nodes/position, WDL enabled, temperature 0.15

Files

Path	Description
checkpoints/formal_1m_latest.pt	Torch checkpoint. Stores architecture (model kwargs), model (state_dict), optimizer/scheduler state, global_step, epoch, and the training config.

How to load

The checkpoint is a project-specific state_dict; you need the project's model code. Clone the repo and download the checkpoint so the path resolves unchanged:

git clone https://github.com/JinShuo-Li/ChessModel.git
cd ChessModel
hf download jinshuoli/chessmodel checkpoints/formal_1m_latest.pt --local-dir .

Load the weights:

import torch
from chess_ai.model import ChessNetwork
from chess_ai.training.checkpoint import load_checkpoint

state = torch.load("checkpoints/formal_1m_latest.pt", map_location="cpu", weights_only=False)
model = ChessNetwork(**state["architecture"])          # 12×192, SE hidden 32
load_checkpoint("checkpoints/formal_1m_latest.pt", model)
model.eval()

Run it as an engine / evaluate it through the project's CLI:

# UCI-style play (batched PUCT search)
python play.py --checkpoint checkpoints/formal_1m_latest.pt --device cuda --simulations 800

# Neural metrics on teacher shards
python evaluate.py --checkpoint checkpoints/formal_1m_latest.pt \
    --dataset data/formal_50k_validation --device cuda

Intended use and limitations

• Intended: chess-engine and reinforcement-learning research, distillation experiments, reproducibility of the project's results.
• Not a ready-to-use HF model: it cannot be run through the standard Inference API or transformers; it requires the project's search/engine code.
• Strength: deliberately modest — this is a compact research model, not a competitive engine. Do not treat it as a strong chess player.

Related

• Dataset: jinshuoli/chessmodel-data
• Source code & documentation: github.com/JinShuo-Li/ChessModel