Apple Silicon (M1/M2/M3 Macs)

• macOS & Xcode: Ensure you are on a recent version of macOS with Xcode and the command line tools installed. No separate GPU drivers are needed, as PyTorch uses the built-in Metal Performance Shaders (MPS) backend.
• Note: While functional, performance and stability on MPS can sometimes lag behind CUDA.

Quick Start

This project uses a Gradio web interface to simplify the fine-tuning process.

1. Installation

First, install the required dependencies:

pip install -r requirements.txt

2. Launch the Web App

Run the app.py script to start the Gradio server:

python3 app.py

This will launch a web server and provide a local URL (usually http://127.0.0.1:7860). Open this URL in your browser.

3. Using the App

The app is organized into three tabs:

1. Setup & Data Processing:

   • Upload your dataset (CSV, Excel, PDF, or Word).
   • Select the Gemma model you want to fine-tune.
   • Enter your Hugging Face token to authenticate.
   • Click Process Data & Prepare for Training.

2. Training:

   • Select the base model for fine-tuning.
   • Check the Run smoke test box for a quick 10-step run to verify the pipeline.
   • Click Start Fine-Tuning and monitor the logs.

3. Inference & Export:

   • Once training is complete, chat with your fine-tuned model in the chat interface.
   • To share your model, enter a repository name (e.g., your-username/my-gemma-finetune) and click Push to Hub.

Repository Layout

├── configs/            # YAML configuration files
├── core/               # Framework source code
│   ├── data.py         # Dataset loading helpers
│   ├── train.py        # LoRA fine-tuning entry point
│   ├── infer.py        # Inference entry point
│   └── utils.py        # Shared utilities
├── cli.py              # Typer-based CLI wrapper
├── requirements.txt    # Python dependencies
└── setup.py            # Installable package definition

License

This project is released under the Apache 2.0 license.