A fine-tuned Mistral 7B model trained on personal email data to generate responses in your unique writing style.
- Personalized Writing Style: Trained on 100,000 personal emails
- Efficient LoRA Fine-tuning: Only 336MB adapter vs 14GB full model
- Web Interface: Simple Flask-based UI for testing
- Docker Support: Easy deployment with GPU acceleration
- Fast Inference: 4-bit quantization for efficient generation
- Base Model: Mistral-7B-v0.1
- Training Method: QLoRA (4-bit quantization + LoRA adapters)
- Dataset: 100,000 personal emails
- Training Time: ~6 hours on RTX 5090
- Final Loss: 0.14
- LoRA Rank: 32
- Learning Rate: 2e-4
- NVIDIA GPU with 16GB+ VRAM
- Docker with NVIDIA Container Toolkit
- CUDA 12.1+
- Clone the repository:
git clone https://github.com/kylefoxaustin/mistral-mail.git
cd mistral-mail-
Download or place your trained model in
models/mistral-editorial-final/ -
Build and run with Docker:
docker build -t mistral-mail .
docker run -it --gpus all -p 8081:8081 mistral-mail- Open http://localhost:8081 in your browser
See TRAINING.md for detailed instructions on fine-tuning Mistral on your own email dataset.
mistral-mail/
├── app.py # Flask web interface
├── train.py # Training script
├── inference.py # Standalone inference
├── requirements.txt # Python dependencies
├── Dockerfile # Container configuration
├── models/ # Model weights (not in repo)
│ └── mistral-editorial-final/
└── data/ # Training data (not in repo)
This project is designed to train on personal email data. Never share your trained model publicly as it may generate text containing personal information from your training data.
Edit config.yaml to adjust:
- Generation parameters (temperature, top_p, etc.)
- Model paths
- Training hyperparameters
MIT License - See LICENSE file
Kyle Fox - Austin, TX
- GitHub: @kylefoxaustin
- Project: MistralMail
- Built with Mistral-7B
- Uses PEFT for LoRA
- Inspired by personal productivity needs
Note: This model is trained on personal data and should not be shared publicly.