parom23/chess_transformer

Chess Transformer — 35M Parameters, Trained from Scratch

Transformer model trained to play chess using reinforcement learning self-play, built entirely from first principles in PyTorch.

Architecture: Custom transformer with chess-specific position encoding
Parameters: 35M
Training: RL self-play with custom reward shaping
Fine-tune: Separate LoRA/PEFT fine-tune of Qwen-0.5B also available

Built as a personal project to learn transformer architecture and RL from the ground up — not a wrapper, not a fine-tune of an existing chess model.

GitHub: https://github.com/paulromanov23/chess_player_final