Jochemvkem/magnusbot
Jochemvkem/magnusbot is a 0.5 billion parameter custom encoder-decoder Transformer model developed by Jochem van Kemenade. Trained from scratch on approximately 4 million chess positions, it specializes in predicting strong chess moves in UCI format given a FEN notation board state. The model was fine-tuned on tactical positions and is designed for integration into chess engines or as a tournament player.
Loading preview...
MagnusBot: A Custom Chess Move Predictor
MagnusBot is a 0.5 billion parameter custom encoder-decoder Transformer model developed by Jochem van Kemenade, specifically designed for predicting optimal chess moves. It takes a board state in FEN notation as input and outputs the predicted best move in UCI format.
Key Capabilities
- Specialized Chess AI: Trained end-to-end on a large dataset of approximately 4 million chess positions, including games and tactical puzzles.
- Two-Phase Training: Underwent a base training phase (25 epochs) followed by a fine-tuning phase (4 epochs) focused on tactical positions, such as checkmate threats and winning combinations.
- Custom Architecture: Utilizes a custom Transformer architecture with a
ChessTokenizertailored for chess-specific vocabulary.
Training Details
The model was trained from scratch using a combination of local datasets and the ssingh22/chess-evaluations tactical puzzles from HuggingFace. Training was performed on an NVIDIA GeForce RTX 4070 Super and took approximately 10 hours.
Use Cases
- Chess Engine Component: Can be integrated into larger chess engines.
- Tournament Player: Suitable for use in automated chess playing systems.
Limitations
- Domain-Specific: This model is highly specialized for chess move prediction and is not intended for general natural language processing tasks. Performance outside its domain will be poor.