jspaulsen/halluci-mate-v1b

Overview

jspaulsen/halluci-mate-v1b is an alpha-release chess Large Language Model (LLM) developed by jspaulsen. It is a fine-tuned continuation of jspaulsen/halluci-mate-v1a, built upon the Qwen3 architecture with approximately 0.6 billion parameters. The model was trained from scratch using the Lichess dataset and features a custom UCI move tokenizer, enabling it to process and generate chess moves.

Key Capabilities

• Chess Move Generation: Predicts chess moves in UCI format based on the current game state.
• Custom Tokenization: Utilizes a specialized tokenizer for geometric and promotion UCI moves, along with special tokens for game conditions (<WHITE>, <BLACK>, <DRAW>).
• Constrained Decoding: Supports masking logits to ensure only legal moves are generated, mitigating illegal-move hallucinations.
• Large Context Window: Features a context length of 32,768 tokens, allowing for analysis of longer game sequences.

Good For

• Experimental Chess AI Development: Ideal for developers exploring LLM applications in chess, particularly for move prediction and game simulation.
• Understanding LLM Behavior in Structured Environments: Useful for studying how LLMs perform in rule-bound systems like chess, especially regarding hallucination control.

Limitations

• Alpha Quality: The model is in an early development stage; move strength and strategic robustness are unvalidated.
• Illegal Move Generation: Without constrained decoding, the raw model may produce illegal move tokens.
• Human Game Bias: Trained on human games, its behavior reflects idiosyncrasies and blunders found in lower-rated play.
• FEN Support: Limited support for analyzing positions from arbitrary FENs beyond what the Game class constructs.