Phi-2 Fine-tuning

A comprehensive toolkit for fine-tuning Microsoft's Phi-2 and Phi-3.5 language models, featuring memory-efficient training, interactive chat, and model comparison capabilities.

🌟 Features

• Fine-tuning with LoRA

  • Memory-efficient training optimized for Apple Silicon (MPS) and CUDA
  • Small adapter files (~20MB vs full model ~2.7GB)
  • Detailed layer control (see PHI2 Layer Guide)

• Interactive Tools

  • dialogue.py: Chat with base or fine-tuned models
  • compare_models.py: Compare base vs fine-tuned responses
  • main.py: Train and create LoRA adapters

📋 Requirements

• macOS 12.3+ with Apple Silicon OR Linux with CUDA
• Python 3.9+
• PyTorch 2.2.0+ (with MPS or CUDA support)
• ~21GB available memory for model comparison

🚀 Quick Start

1. Installation

git clone https://github.com/lpalbou/phi2-finetuning.git
cd phi2-finetuning
python -m venv .venv
source .venv/bin/activate  # On Unix/macOS
pip install -r requirements.txt

2. Fine-tune a Model

python src/main.py \
    --output_dir ./output/my_model \
    --dataset_path ./data/my_dataset.jsonl

3. Chat with Models

# Use base model
python src/dialogue.py

# Use fine-tuned model
python src/dialogue.py --adapter_path output/my_model/final_adapter

4. Compare Models

python src/compare_models.py --adapter_path output/my_model/final_adapter

📁 Project Structure

.
├── src/
│   ├── config/            # Training configurations
│   ├── trainers/          # Training implementations
│   ├── callbacks/         # Training callbacks
│   ├── utils/             # Utility functions
│   ├── main.py            # Training entry point
│   ├── dialogue.py        # Interactive chat
│   └── compare_models.py  # Model comparison
├── data/                  # Training datasets
└── docs/                  # Additional documentation

🛠️ Tools Guide

Training (main.py)

The primary tool for fine-tuning models. See our Training Guide for detailed instructions and parameters.

Chat (dialogue.py)

Interactive chat interface supporting:

• Base model (Phi-2 or others)
• Fine-tuned model (base + LoRA adapter)
• Custom prompts and parameters

Compare (compare_models.py)

Side-by-side comparison tool to evaluate fine-tuning effects:

• Interactive REPL mode for live testing
• Batch mode with YAML question files
• Visual output with colored responses

📚 Documentation

• Training Guide - Comprehensive training documentation
• PHI2 Layer Guide - Detailed model layer explanations
• Examples - Sample datasets and configurations

🤝 Contributing

1. Fork the repository
2. Create your feature branch
3. Commit your changes
4. Push to the branch
5. Open a Pull Request

📜 License

MIT License

🙏 Acknowledgments

• Microsoft for the Phi-2 model
• Hugging Face for the transformers library
• PEFT library for LoRA implementation