Raj-Patel7807/rooksac-policy-value

RookSac Policy Value Network v1

The neural network model used by RookSac, a C++20 chess engine.

Architecture

• Architecture: PolicyValueNet-ResNet
• Input representation: 18 × 8 × 8
• Piece planes: 12
• Side-to-move planes: 1
• Castling-rights planes: 4
• En-passant planes: 1
• Policy vocabulary size: 4240
• Schema version: 2.0.0
• Vocabulary version: 2.0.0
• Value perspective: White, in [-1, 1]

Training

• Framework: PyTorch
• GPU: NVIDIA Tesla T4
• Dataset: Lichess standard chess PGN archive
• Games processed: 2,000
• Minimum player rating: 1800
• Epochs: 1
• Batch size: 32
• Learning rate: 0.001
• Residual blocks: 2
• Channels: 32
• Random seed: 42

Evaluation

Metric	Validation	Test
Policy Top-1 Accuracy	14.44%	14.04%
Policy Top-3 Accuracy	29.75%	28.68%
Value MSE	0.98995	0.98598
Value MAE	0.91784	0.91823

Dataset and licensing

Training data was extracted from Lichess standard chess PGN archives.

Dataset source: https://database.lichess.org/

Dataset licensing and applicable terms must be reviewed separately. The model publisher does not claim ownership of the original game records.

Limitations

• This is an initial experimental model.
• It was trained on a relatively small dataset.
• One training epoch is insufficient to establish strong chess-playing ability.
• Playing strength has not yet been established through engine-vs-engine testing.

Files

• model.rks: Native RookSac model artifact.
• training_summary.json: Training configuration and evaluation metrics.
• training_metrics.csv: Epoch-level training metrics.
• model.sha256: SHA-256 checksum of the native model artifact.