Supervised Fine-Tuning (SFT) uses GameLab’s human gameplay data to train models on high-quality examples of real decision-making. By learning directly from human actions in well-defined contexts, models develop strong baseline behaviors that reflect effective strategies and practical reasoning patterns.

This step plays a critical role in bootstrapping model performance before reinforcement learning. It provides a stable starting point that reduces exploration inefficiencies and improves convergence, enabling models to build on human-like behaviors as they transition into more advanced learning phases such as RL and self-play.