#Chess Engine

Custom chess neural network trained on high-elo Lichess games.

Architecture

  • Backbone: ResNet with 10 blocks, 128 channels
  • Parameters: 44.98M
  • Policy Head: 20480 move logits
  • Value Head: Win probability [-1, 1]

Training

  • Phase 1: Supervised learning on 1800+ Elo games
  • Phase 2: Self-play reinforcement learning

Board Encoding

  • 19 input planes (12 piece planes + side to move + castling + en passant + move clock)
  • 8x8 board representation

Move Encoding

  • Flat index: (from_square * 64 + to_square) * 5 + promotion_type
  • Total vocabulary: 20480 moves
Downloads last month
73
Safetensors
Model size
45M params
Tensor type
F32
·
Video Preview
loading