🚀 FineTune

A complete, beginner-friendly system for fine-tuning models compatible with Ollama using datasets in JSON, CSV, or Excel format.

✨ Features

• 🤖 Model Selection: Choose from popular models (LLaMA2, Mistral, Phi, etc.)
• 📁 Multi-Format Support: Upload JSON, CSV, or Excel datasets
• ⚙️ Easy Configuration: Interactive column mapping and parameter tuning
• 🔥 Efficient Training: Uses Unsloth for fast, memory-efficient fine-tuning
• 🚀 Ollama Integration: Automatic model deployment to Ollama
• 🖥️ Dual Interface: Both CLI and web-based interfaces
• 🧪 Built-in Testing: Test your models with custom prompts

🚀 Quick Start

Prerequisites

1. Install Ollama (if not already installed):

   curl -fsSL https://ollama.com/install.sh | sh
   ollama serve  # Start Ollama server

2. Clone and setup FineTune:

   git clone <your-repo-url>
   cd FineTune
   pip install -r requirements.txt

Option 1: Web Interface (Recommended for Beginners)

streamlit run finetune_app.py

Then open your browser to http://localhost:8501 and follow the guided interface!

Option 2: Command Line Interface

python finetune_cli.py

Follow the interactive prompts to:

1. Select your dataset
2. Configure column mappings
3. Choose a model
4. Start training
5. Deploy to Ollama

📖 Step-by-Step Guide

1. Prepare Your Dataset

Your dataset can be in any of these formats:

JSON Format

[
  {
    "question": "What is AI?",
    "answer": "AI stands for Artificial Intelligence..."
  },
  {
    "question": "How does ML work?", 
    "answer": "Machine Learning works by..."
  }
]

CSV Format

prompt,response
"Hello, how are you?","I'm doing well, thank you!"
"What's the weather like?","I don't have access to current weather data..."

Excel Format

Create an Excel file with columns like:

• Task | Context | Expected_Output
• Input | Target
• Question | Answer

2. Select Your Model

Choose from these pre-configured models:

• LLaMA 2 7B - General purpose, good balance
• Mistral 7B - Fast and efficient
• Phi 3 Mini - Lightweight, great for beginners
• Code Llama 7B - Specialized for code
• Gemma 7B - Google's efficient model

3. Configure Training

Basic Settings (good defaults provided):

• Max steps: 60 (adjust based on dataset size)
• Learning rate: 2e-4
• Batch size: 2 (increase if you have more GPU memory)

Advanced Settings:

• LoRA rank: 16 (higher = more parameters, longer training)
• Sequence length: 2048 tokens

4. Train Your Model

Training will:

• Process your dataset into the correct format
• Load the model with LoRA adapters
• Train efficiently using Unsloth
• Save the fine-tuned model

Training Time Estimates:

• Small dataset (100 examples): ~5-10 minutes
• Medium dataset (1000 examples): ~30-60 minutes
• Large dataset (10000 examples): ~2-5 hours

5. Deploy to Ollama

The system will:

• Generate a Modelfile for Ollama
• Register your model with Ollama
• Make it available for chat/API use

6. Test Your Model

# Test with CLI
python test_model.py --model your-model-name --interactive

# Test with Ollama directly
ollama run your-model-name "Hello, how are you?"

📊 Example Datasets

The project includes sample datasets in the datasets/ folder:

• sample_qa.json - Question & Answer pairs
• sample_chat.csv - Conversational responses
• sample_instruction.xlsx - Task-based instructions

Use these to test the system before using your own data!

🛠️ Advanced Usage

CLI Options

# Interactive mode
python finetune_cli.py

# Test a deployed model
python test_model.py --model custom-llama2 --prompt "Hello!"

# Compare multiple models
python test_model.py --compare model1 model2 --file test_prompts.json

# Benchmark performance
python test_model.py --model custom-llama2 --benchmark

Configuration

Edit config.yaml to customize:

• Default training parameters
• Supported models list
• File paths
• Ollama settings

Custom Templates

For multi-column inputs, use templates:

"Question: {question} Context: {context}"

This combines multiple columns into a single input.

🔧 Troubleshooting

Common Issues

1. CUDA Out of Memory

• Reduce batch size to 1
• Enable 4-bit loading
• Use a smaller model (Phi 3 Mini)

2. Ollama Connection Failed

• Make sure Ollama is running: ollama serve
• Check if port 11434 is available
• Try restarting Ollama

3. Training Too Slow

• Use fewer training steps
• Increase batch size (if memory allows)
• Try gradient accumulation

4. Poor Model Performance

• Increase training steps
• Use more/better training data
• Try different learning rates
• Ensure data quality

Getting Help

1. Check the example datasets for formatting
2. Review the configuration in config.yaml
3. Use the web interface for guided setup
4. Test with small datasets first

📁 Project Structure

FineTune/
├── src/                     # Core modules
│   ├── config.py           # Configuration management
│   ├── dataset_processor.py # Dataset handling
│   ├── model_manager.py    # Model loading/management
│   ├── trainer.py          # Training engine (Unsloth)
│   └── ollama_integration.py # Ollama deployment
├── datasets/               # Sample datasets
├── models/                 # Downloaded models
├── outputs/                # Training outputs
├── config.yaml            # Main configuration
├── finetune_app.py        # Streamlit web interface
├── finetune_cli.py        # Command line interface
├── test_model.py          # Model testing script
├── test_prompts.json      # Sample test prompts
├── requirements.txt       # Dependencies
└── README.md              # This file

🤝 Contributing

Contributions welcome! Please:

1. Fork the repository
2. Create a feature branch
3. Make your changes
4. Test thoroughly
5. Submit a pull request

📄 License

This project is open source and available under the MIT License.

🙏 Acknowledgments

• Unsloth - For efficient fine-tuning
• Ollama - For local model deployment
• Hugging Face - For model ecosystem
• Streamlit - For the web interface

---

Happy Fine-tuning! 🎉

Need help? Check the examples, try the web interface, or create an issue on GitHub.