Google Tunix Hack - Train a model to show its work

Overview

This hackathon challenges you to fine-tune Google's open-weight Gemma models using Tunix, a new JAX-native library, to teach them how to reason through complex questions and 'show their work'. You'll create a model that not only provides the correct answer but also explains the steps it took to get there, making AI more capable, trustworthy, and transparent.

Requirements

You need to create a working training pipeline using Tunix and either the Gemma2 2B or Gemma3 1B model. Participants must share their configurations, reward function composition, and recipes so others can reproduce and build upon their work. Your submission needs to include a Kaggle Writeup with a title, subtitle, and detailed analysis (under 1,500 words). This Writeup must have a cover image in the Media Gallery, a public Kaggle Notebook with your code, and a YouTube video (3 minutes or less) explaining your project. The fine-tuned model should be directly output from the notebook and runnable on a single Kaggle TPU session, producing output in a specific format: 'modelthinkingtracemodel_answer'.

Prizes

There's a total of $100,000 in prizes across the main track. The top six places will receive cash awards: 1st place gets $30,000, 2nd place receives $25,000, 3rd place gets $15,000, and 4th, 5th, and 6th place each receive $10,000.