Agentic AI
Community Developers Train Gemma Models for Reasoning with Tunix and TPUs
Over 11,000 developers competed to build reasoning capabilities into small language models using Google's Tunix framework and free Kaggle TPUs, with winning teams demonstrating sophisticated post-training techniques.
Google Developers Blog (AI)
Google challenged the developer community to transform non-reasoning base models into general reasoning models using Tunix and Kaggle TPUs. The Google Tunix Hack hackathon saw overwhelming participation: over 11,000 entrants submitted more than 300 high-quality entries, proving that quality reasoning training is achievable on a budget—using just a Kaggle TPU v5e-8 for 9 hours.
Winning Approaches#
The top solutions combined supervised learning, preference optimization, and reinforcement learning in creative ways:
G-RaR (Rubrics as Rewards) trained Gemma-2-2B to produce structured reasoning by combining Supervised Fine-Tuning (SFT) with GRPO, using a rubric-based LLM-as-judge reward system. The model learns to "show its work" inside <reasoning> tags before outputting answers. A larger Gemma-3-12B judge model evaluates intermediate logical steps, converting discrete rubric scores into continuous reward signals. A two-stage pipeline (SFT warm-start, then GRPO refinement) used a split-mesh architecture on a single TPU to parallelize policy and judge models.
"Pinocchio" evolved a 1B parameter model into a reasoning engine through a three-stage, 9-hour pipeline: SFT with distillation from an OSS-120B teacher, SimPO for strict formatting, and GRPO refinement using Gemini 2.0 Flash as an asynchronous judge. The team extended Tunix itself, injecting a custom SimPO loss function with length normalization and building a high-throughput evaluation engine.
IDEA-E Framework distilled structured ethical reasoning into a 2B model using curriculum-guided GRPO and a fast TF-IDF reward system. Rather than relying on slow LLM judges, TF-IDF incentivizes context-relevant vocabulary in reasoning traces, preventing verbose outputs and enabling rapid, non-blocking calculations on the CPU.
Domain-Specific Wins#
Several submissions showcased reasoning training tailored to specific fields:
- Medical: Structured reasoning traces enhanced interpretability in clinical problem-solving
- Chemistry: Step-by-step traces helped small models tackle complex chemistry reasoning
- Legal: GRPO post-training enabled accurate analysis of complex legal data
- Robotics: Multi-step reasoning generation solved planning and decision-making tasks
