huggingface/huggingface-gemma-recipes

What it solves

This repository provides minimal, practical recipes for working with the Gemma family of models. It lowers the barrier for developers to implement multimodal inference, fine-tuning, and retrieval-augmented generation (RAG) using these models.

How it works

The project consists of a collection of Colab notebooks and Python scripts that demonstrate how to use the transformers and timm libraries to interact with Gemma models. It provides concrete code examples for:

• Inference: Running text, image, audio, and video inputs through the model using pipelines or detailed processor configurations.
• Fine-tuning: Using tools like Unsloth and TRL to adapt models for conversational tasks, GUI grounding, or specific modalities like radiology images or audio.
• RAG: Implementing retrieval-augmented generation workflows.
• Function Calling: Demonstrating how to use Gemma 3n for tasks like local file reading.

Who it’s for

Developers and AI researchers who want to quickly deploy or customize the Gemma family of models for multimodal applications.

Highlights

• Multimodal Support: Recipes for handling interleaved text, image, audio, and video inputs.
• Diverse Fine-tuning: Examples for conversational tuning, GUI grounding, and multimodal fine-tuning on specific datasets.
• Hardware Accessibility: Notebooks optimized for free Colab T4 GPUs.
• Advanced Techniques: Includes examples for Reinforcement Learning (GRPO) and function calling.