Multi-Domain GRPO Reasoning

Google Tunix Hackathon project teaching Gemma2 to reason transparently with multi-domain GRPO training.

Multi-Domain GRPO Reasoning explores transparent reasoning training for Gemma2 across multiple domains.

  • Fine-tuned Gemma2-2B-IT across 5 domains (GSM8K, SciQ, CNN/DailyMail, synthetic creative, MBPP) with Google’s Tunix GRPO library.
  • Improved format accuracy 40.8% → 100% and exact accuracy 16.5% → 60.8%.
  • Implemented GRPO from scratch: rollouts, group-relative advantages, KL-regularized updates, and AST-based validators.
  • Surprise finding: science (70%) beat math (58%) despite 3× less data — evidence of transferable reasoning.

View the repository