KataGo x LLM Explainable AI
Fine-tuning LLMs to explain Go strategies using serialized game states.
| Role: Project Lead | Stack: Python, Qwen2.5, Hugging Face, KataGo |
[cite_start]As the Project Lead for this initiative, I am coordinating a team of 6 contributors to build an Explainable AI system for the game of Go[cite: 35, 40]. The goal is to fine-tune Large Language Models (LLMs) to provide natural language explanations for complex game situations.
Key Technical Contributions:
- Custom GTP Proxy: Wrote
my_gtp_proxy.pyto intercept Go Text Protocol (GTP) commands from Lizzie/KataGo. [cite_start]This allows us to serialize game states and KataGo’s analysis into JSONL format in real-time[cite: 36]. - Data Pipeline: Designed a pipeline to parse SGF records and query KataGo-18B. [cite_start]We extract top-k candidate moves along with win-rates and score-leads to create a filtered dataset for reward signals[cite: 37].
- LLM Fine-tuning: Adapted Hugging Face scripts to fine-tune Qwen2.5-7B-Instruct using LORA/GRPO techniques. [cite_start]I also set up local inference prototyping using LM Studio[cite: 39].