Instead of a standard hackathon where the focus is often on rapid product prototyping, the Gemma Fine-Tuning Challenge shifts the spotlight directly onto core machine learning engineering.
Your mission is to build a three-way claim-verification classifier. Using a curated synthetic dataset, you will fine-tune a Gemma model to analyze a claim alongside evidence passages and predict exactly one label: SUPPORTS, REFUTES, or NOT_ENOUGH_INFO. You'll need to optimize QLoRA hyperparameters, manage computational efficiency, and architect a robust data pipeline to squeeze every ounce of performance out of the 2B baseline model.
Gemma is a family of lightweight, state-of-the-art open models built from the same research and technology used to create the Gemini models. By fine-tuning Gemma, you get hands-on experience with modern LLM architectures that can run anywhere.
Our judges (Google Developer Experts and ML Researchers) will evaluate your submission based on three core technical pillars.
The ultimate test of your model's capacity. You will be evaluated on a hidden test set based on your Exact Accuracy and Macro-F1 score. Generalization is key to winning this category.
You are restricted to a concise 3-slide presentation. Our judging panel will evaluate your validation accuracy, prompt design, technical strategy, and ability to articulate your learnings.
You will be provided a Colab notebook implementing a Gemma baseline. You are permitted to use any Gemma model, tweak the baseline, or create your own, provided you follow our data standards.
Any use of a non-Gemma based model will result in disqualification.