mattmireles

mattmireles / gemma-tuner-multimodal

Public

Fine-tune Gemma 4 and 3n with audio, images and text on Apple Silicon, using PyTorch and Metal Performance Shaders.

736
35
100% credibility
Found Apr 09, 2026 at 736 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

A user-friendly tool for customizing Google's Gemma AI models with personal text, image, or audio data directly on Apple Silicon Macs, featuring an interactive wizard and real-time training visualization.

How It Works

1
🔍 Discover easy AI customization

You hear about a simple tool that lets you teach Google's Gemma AI to understand your own pictures, sounds, or text right on your Mac.

2
📦 Quick setup

Follow a few easy steps to install it, like adding a helper app that works perfectly on your Apple computer.

3
🧙 Start the friendly guide

Open the magic wizard that asks simple questions to pick a smart AI brain and your personal data files.

4
🚀 Train your custom AI

Hit go and watch your AI learn from your examples, with a live colorful view in your web browser showing progress.

5
📱 Test and improve

Try your trained AI on new pictures or sounds, see how well it understands, and tweak if needed.

🎉 Your personal AI is ready

Celebrate having a custom AI that perfectly handles your specific images, voices, or text, all running smoothly on your Mac.

Sign up to see the full architecture

4 more

Sign Up Free

Star Growth

See how this repo grew from 736 to 736 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is gemma-tuner-multimodal?

This Python tool lets you fine tune Gemma locally with LoRA on Apple Silicon using PyTorch and Metal Performance Shaders—no NVIDIA GPU needed. It handles text-only instruction tuning, image captioning/VQA, or audio transcription from CSV datasets, even streaming terabytes from GCS or BigQuery without local storage. Run `gemma-macos-tuner wizard` for guided setup or `finetune ` to train Gemma 3n/4 models like gemma-3n-E2B-it on your Mac.

Why is it gaining traction?

Unlike MLX-LM or Unsloth, it supports full multimodal fine tune Gemma 3n/4b with audio/images/text natively on MPS, plus a live browser visualizer showing loss curves, attention heatmaps, and memory in real time. Developers skip H100 rentals and data copying for domain-specific VLM or ASR adaptation. The CLI wizard auto-generates configs, and exports merge LoRA into HF/SafeTensors for Hugging Face or Ollama.

Who should use this?

ML engineers adapting Gemma for private on-device ASR (medical/legal audio) or vision tasks (receipts/charts) on M1/M2/M3 Macs. Teams fine tuning VLMs with custom CSV without cloud infra, or researchers prototyping gemma 3 270m/4b LoRA on terabyte datasets via BigQuery streaming.

Verdict

Solid for Apple Silicon users wanting quick multimodal fine tune Gemma 3—736 stars show interest, docs are thorough with guides. But 1.0% credibility score flags early maturity; test small before production.

(198 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.