Minis

Overview

Minis generates personalized micro-models from a developer's GitHub profile. Point it at a GitHub user, it scrapes their public repos and commit history, and fine-tunes a small language model via LoRA to capture their coding patterns, documentation style, and technical vocabulary.

The output: a lightweight neural artifact — a LoRA adapter — that can be loaded on top of a base model to produce outputs "in the voice" of that developer.

Problem

Fine-tuning large language models is expensive, slow, and requires significant data curation. Most personalization approaches assume you have a clean, curated dataset. GitHub profiles are the opposite: messy, varied, multilingual across programming languages — but rich with authentic developer signal.

The question: can you distill a developer's unique style into something small enough to run locally?

Solution

Minis is a Python pipeline built on PyTorch, Hugging Face PEFT, and the GitHub API:

• Profile scraper — fetches READMEs, comments, docstrings, commit messages, and inline code from public repos
• Data preprocessor — filters noise, normalizes formatting, and builds a training corpus
• LoRA trainer — fine-tunes a small base model (e.g., Phi-2, Mistral 7B Q4) using PEFT/LoRA for parameter-efficient adaptation
• Adapter packaging — exports a .safetensors LoRA adapter + metadata for re-use or sharing
• Inference CLI — test your mini-model locally with minis generate --user <github_handle>

Training a mini takes minutes on a consumer GPU, not hours — LoRA keeps the parameter count tiny while still capturing meaningful stylistic signal.

Challenges

GitHub is not a clean dataset. Boilerplate, generated code, and dependency lock files dominate many repos. Getting the signal-to-noise ratio right required aggressive filtering and tokenization heuristics. LoRA rank selection also matters a lot: too low and nothing sticks, too high and you overfit to a handful of files.

Impact

• Demonstrates practical LoRA fine-tuning on unstructured, real-world data
• Opens the door to "developer digital twins" — models that can continue someone's work in their style
• Useful as a code-style transfer tool, documentation generator, or creative exploration of identity in AI

Reflection

Minis started as a weekend experiment: "can a model learn me from just my public GitHub?" The answer is a qualified yes — the adapter picks up vocabulary, commenting habits, and structural preferences reliably. The deeper question of what that means for identity, authorship, and AI attribution is still wide open, and that's what makes it interesting.

Tech Stack

Python, PyTorch, Hugging Face Transformers, PEFT / LoRA, GitHub REST API, safetensors, Mistral / Phi-2 base models