Fine-tuning a 350M Model for Better Structured Outputs in 100 GRPO Steps