How the Community Trained Gemma to "Think" with Tunix and TPUs (developers.googleblog.com)

In a recent initiative, developers transformed non-reasoning models into general reasoning engines through the Google Tunix Hack. This hackathon attracted over 11,000 participants who used Kaggle TPUs to train the Gemma models (Gemma-2-2B and Gemma-3-1B) to produce structured reasoning traces, significantly advancing the community's capability to enhance large language models. The winning approaches utilized innovative techniques like Supervised Fine-Tuning (SFT), preference optimization via GRPO, and reinforcement learning, effectively demonstrating that high-quality reasoning can be achieved with limited computational resources. The significance of this event lies in its role in democratizing AI development by providing accessible training recipes that include data, strategies, and runnable code. The advancements made in post-training pipelines allowed models to generate coherent reasoning outputs, shifting from simple pattern matching to logical deduction across various domains such as medicine, chemistry, legal frameworks, and robotics. Noteworthy innovations included the use of rubric-based reward systems and the IDEA-E ethical reasoning framework, which facilitates logical deductions by enforcing structured reasoning steps. Overall, this hackathon has produced valuable resources for the AI/ML community, enabling developers to create robust reasoning models on easily accessible hardware.