SpinGPT: Large-Language-Model Approach to Poker

This technical report presents SpinGPT, a large-language-model (LLM) designed specifically for playing Spin & Go, a three-player online poker format. The report outlines the limitations of the Counterfactual Regret Minimization (CFR) algorithm in multi-player settings and discusses the two-stage training process employed for SpinGPT. The first stage involves Supervised Fine-Tuning (SFT) on a dataset of 320,000 high-stakes poker decisions, while the second stage utilizes Reinforcement Learning (RL) with a dataset of 270,000 solver-generated hands to refine the model's strategy towards game-theoretic optimality. The report details the model architecture based on Llama-3.1-8B-Instruct and describes the data preprocessing steps taken to prepare the dataset for training. Results indicate that SpinGPT matches solver actions in 78% of decisions, demonstrating its potential effectiveness in multi-player imperfect-information games like poker. The findings suggest that LLMs could offer a new approach to tackling the complexities of poker strategy.