llm-chess-agent-v3-half-training

LLM Chess Agent - Global Chess Challenge 2025

This model is a fine-tuned chess agent for the Global Chess Challenge 2025.

🎯 Architecture

• Base Model: Qwen/Qwen2.5-1.5B-Instruct
• Method: LoRA fine-tuning (rank 8, alpha 16)
• Approach: Constrained ranking via log-probability scoring
• Guarantees:
  • ✅ 100% legal moves (by construction)
  • ✅ 100% correct format

🎮 How It Works

The agent uses constrained ranking instead of free generation:

1. Environment provides: FEN + side + list of legal moves
2. Agent scores each candidate move via log-probability
3. Agent selects: best_move = argmax(scores)
4. Result: Always legal (move is always in the provided list)

📊 Performance

• Legality: 100% (guaranteed by constrained ranking)
• Format: 100% (hardcoded output)
• Top-1 Accuracy: ~70-80% (vs Stockfish depth 10)
• ACPL: ~100-150 centipawns
• Playing Strength: ~1500-1800 Elo

🚀 Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

# Load base model
base_model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-1.5B-Instruct")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-1.5B-Instruct")

# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, "SBellilty/llm-chess-agent-v3-half-training")

# Use with the official challenge environment
# See: https://github.com/AIcrowd/global-chess-challenge-2025-starter-kit

📝 Training

• Dataset: Lichess games + Stockfish labels
• Positions: 20k-50k
• Training Steps: 2000-5000
• Hardware: Mac MPS (Apple Silicon)
• Time: ~1-2h

🏆 Challenge

Submitted to the Global Chess Challenge 2025:

• https://www.aicrowd.com/challenges/global-chess-challenge-2025

📄 License

MIT License

🙏 Acknowledgments

• Challenge organizers: AIcrowd & AGI House
• Base model: Qwen team
• Chess engine: Stockfish
• Data source: Lichess Open Database