by Mauro Serralvo, founder at BrinPage.

BirdMind v0 is our very first prototype — a sandbox experiment designed to test the training pipeline, fine-tuning workflow, and deployment setup for a code generation model. This version is not intended for production use, but rather as a learning step and foundation for future development.

The production-ready model, BirdMind v1, is introduced on the main product page. That version will be trained on curated datasets and optimized for real-world deployments. The work documented here represents the groundwork leading up to it.

For BirdMind v0, we start with CodeParrot as a base model and perform fine-tuning. The goal is to validate the process step by step: environment setup, dataset preparation, training, and inference.

BirdMind v0 will not only include the model itself, but also a minimal interface built with Next.js and an API endpoint that mirrors the OpenAI format, allowing developers to interact with the model in a simple and standardized way. This mirrors the workflow of real-world deployments, where both inference and integration matter as much as training.

To begin, we first set up the development environment:
• Install Python 3.12.8
• Install recommended VS Code extensions:autopep8 (for formatting) and isort (for imports)
• Create a virtual environment with venv using Python 3.12.8

For the model, we will use CodeParrot-Small from Hugging Face as the base, and apply fine-tuning on top of it.