CodeAnvil 🔨

Fine-tune code language models on your codebase to create specialized AI coding assistants.

What is CodeAnvil?

CodeAnvil automates the process of fine-tuning large language models (like DeepSeek Coder) on your specific codebase. The result is an AI assistant that deeply understands your project's architecture, coding patterns, and domain-specific knowledge.

Quick Start

Requirements

• OS: Rocky Linux 10, Arch Linux, or similar
• GPU: NVIDIA GPU with 16GB+ VRAM (RTX 4080 recommended)
• RAM: 16GB minimum, 32GB+ recommended
• Disk: 100GB free space
• CUDA: 12.1 or later

Installation

git clone https://github.com/mdmattsson/codeanvil.git

-- or --

# Download the script
wget https://codeanvil.net/codeanvil.sh
chmod +x codeanvil.sh


# Run setup (installs all dependencies)
./codeanvil.sh setup

# Add alias to your shell
echo 'alias codeanvil="${HOME}/codeanvil.sh"' >> ~/.bashrc
source ~/.bashrc

Basic Usage

# Interactive menu (easiest)
codeanvil

# Train a model on your codebase
codeanvil train /path/to/myproject

# Monitor training progress (in another terminal)
codeanvil monitor

# Export trained model
codeanvil export deepseek-coder

# Test your model
ollama run codeanvil-deepseek_coder:v1.0

# Use with Aider for AI pair programming
aider --model ollama/codeanvil-deepseek_coder:v1.0

What Gets Installed

• Python 3.9+ with PyTorch and CUDA support
• LLaMA Factory (training framework)
• Ollama (model serving)
• DeepSeek Coder base model
• All dependencies and tools

Training typically takes 2-4 hours on an RTX 4080 with a typical codebase.

Documentation

• Architecture Guide - How the script works and where files are stored
• Development Guide - How to modify and extend the script
• Training Technical Guide - Deep dive into the training process

Features

• ✅ One-command setup - Automated installation of all dependencies
• ✅ Interactive training - Simple menu-driven interface
• ✅ Real-time monitoring - Watch training progress with live metrics
• ✅ Automated data preparation - Extracts and formats code from your codebase
• ✅ Model export - One-command export to Ollama for easy deployment
• ✅ Multi-model support - Train DeepSeek Coder, Qwen, CodeLlama, and more
• ✅ Comprehensive training - Learns build systems, classes, architecture, and code patterns

Use Cases

AI Pair Programming

# Train on your codebase
codeanvil train /path/to/myproject

# Use with Aider for code editing
aider --model ollama/codeanvil-deepseek_coder:v1.0

Codebase Q&A

# Ask questions about your code
ollama run codeanvil-deepseek_coder:v1.0

>>> How do I build this project?
>>> What does the ImageProcessor class do?
>>> Show me the DICOM processing workflow

Code Review & Refactoring

Use your trained model to understand legacy code, suggest improvements, and maintain consistency with your coding patterns.

Training Configuration

CodeAnvil uses optimized settings for effective fine-tuning:

• Method: LoRA (Low-Rank Adaptation)
• Rank: 128 (high capacity for learning)
• Learning Rate: 5e-4 (aggressive to override base knowledge)
• Epochs: 10 (sufficient repetition)
• Quantization: 4-bit (fits in 16GB VRAM)

These settings are tuned to help the model learn your specific codebase while maintaining general coding knowledge.

Support

• Website: codeanvil.net
• Issues: Open an issue on GitHub
• Documentation: See docs/ folder

License

MIT License - See LICENSE file for details

---

CodeAnvil - Forge custom AI models from your code 🔨