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

The blog details an end-to-end GRPO training pipeline to teach Gemma-3 to solve GSM8K math problems. The pipeline includes environment setup, Hugging Face authentication, loading Gemma-3, and wrapping examples into a 'reasoning plus answer' prompt format. It then defines reward functions for format compliance and numerical correctness, attaches LoRA adapters for lightweight training. Finally, it evaluates the base model, generates group samples for GRPO to improve policy, and optionally exports the merged model. For developers or research teams wanting to fine-tune Gemma-3 for math reasoning, this workflow provides a low-resource implementation. Using LoRA adapters allows training on a single consumer-grade GPU without large clusters. Combining GRPO with GSM8K rewards effectively boosts the model's reasoning on structured math problems.