UofTCSSLab/C1-4B
UofTCSSLab/C1-4B is a 4 billion parameter language model developed by UofTCSSLab, specifically fine-tuned for grounded chess reasoning. This model analyzes chess positions step-by-step in natural language and outputs the best move in UCI notation. It achieves 48.3% accuracy on a 900-puzzle test set, demonstrating its specialization in strategic chess analysis. The model is designed for applications requiring detailed, human-readable chess move explanations and optimal move suggestions.
Loading preview...
UofTCSSLab/C1-4B: Grounded Chess Reasoning Model
UofTCSSLab/C1-4B is a 4 billion parameter language model developed by UofTCSSLab, specifically engineered for advanced chess reasoning. This model excels at analyzing complex chess positions, providing step-by-step natural language explanations, and identifying the optimal move in UCI notation.
Key Capabilities
- Grounded Chess Reasoning: Analyzes chess positions (provided in FEN, piece arrangements, and legal moves) and generates detailed reasoning.
- Optimal Move Prediction: Concludes its analysis with a single best move in UCI format (e.g.,
e2e4). - Performance: Achieves a 48.3% accuracy on a 900-puzzle test set for exact-match UCI moves, outperforming its SFT-stage base model (42.3%).
- Instruction-Tuned: The model is the final (SFT + RL) version, indicating refinement through supervised fine-tuning and reinforcement learning.
Recommended Usage
- Greedy Decoding: It is strongly recommended to use greedy decoding (temperature 0) for optimal performance and consistent output.
- Prompt Structure: Users should provide the FEN, piece positions, and legal moves, requesting a step-by-step analysis ending with
FINAL_ANSWER: <uci_move>.
Good for
- Chess Analysis Tools: Integrating AI for explaining chess moves and strategies.
- Educational Platforms: Providing detailed insights into chess game progression.
- Automated Chess Commentaries: Generating human-like commentary for chess matches.