Training Gemma-3 for Structured Mathematical Reasoning with Tunix GRPO, LoRA Adapters, and GSM8K Rewards

We build an end-to-end GRPO training workflow that teaches Gemma-3 to reason through GSM8K math problems. We prepare the environment, authenticate with Hugging Face, load Gemma-3, and wrap examples into a reasoning-plus-answer prompt format.

We define reward functions for format adherence and numeric correctness, then attach LoRA adapters to keep training lightweight. We evaluate a baseline, run GRPO to improve the policy through group-sampled generations, and optionally export the merged model.