engdarwish/chess-position-evaluator

Chess Position Evaluator - V7 HULK

Ahmed Darwish | @engdarwish

Architecture

SE-ResNet-20 with Dual Head (Value + Policy)

Metric	Value
Parameters	47,450,753
Best Val Loss	0.030174
Architecture	SE-ResNet-20
Input	18x8x8 board tensor
Value Output	[-1, 1] (tanh)
Policy Output	20,480 move logits (from-square x to-square x underpromotion)

Training Data

Source	Size
Kaggle Chess Evaluations	12.9M positions (Stockfish evals)
Lichess 2023-10	2M moves from ELO 2000+ games

Quick Start

pip install torch python-chess huggingface_hub numpy

from inference import load_model, get_best_move
import chess

model = load_model()
value, moves = get_best_move(model, chess.STARTING_FEN, top_k=3)
print(value, moves)

See inference.py in this repo for the full, runnable example: board encoding, model definition, and legal-move-masked move ranking.

Files

File	Purpose
model_weights.pt (190MB)	Inference weights: use this to load the model
checkpoint.pt (569MB)	Full training checkpoint (model + optimizer + scheduler state): only needed to resume training
config.json	Architecture config
inference.py	Minimal working inference example
training_curves.png	Training/validation loss curves

Contact

eahmeddarwish@gmail.com GitHub