What is llm-tuning?

This repository contains code for running and customizing two families of large AI language models: Google's Gemma 4 and Alibaba's Qwen3.6. The whole thing is designed to deploy on Modal, a cloud service that rents out GPU machines by the minute. You pay only while your model is running, and Modal handles the infrastructure so you do not need to manage servers yourself. The repo is split into three self-contained sections. The first two (one for Gemma 4, one for Qwen) let you start a model and give it an endpoint that other programs can talk to using the same format as the OpenAI API. That means any tool already built to work with ChatGPT can be pointed at your self-hosted model with a URL change. Each model can be configured two ways: solo mode reserves the full GPU for one person with a very large working memory, and concurrent mode shares the GPU among several users at once with a smaller window each. The third section is a research pipeline showing how to take a model and teach it to be better at a specific task. The example task is answering questions by querying a SQL database. The pipeline runs through four steps: first serve the model, then score how well it performs, then use a larger model to generate training examples, and finally run a targeted fine-tuning pass that teaches the smaller model from those examples without damaging its existing abilities. The repo also includes custom chat templates for both model families. Chat templates are the formatting rules that tell the model how a conversation is structured: where a user message begins, where a tool response goes, and so on. The ones here fix several bugs that cause problems when AI coding tools send multi-step requests with tools attached. All model weights used here are released under Apache 2.0 and do not require any account or token to download. The code itself is also Apache 2.0. Setup requires a Modal account, a Python package manager called uv, and a machine with a GPU if you want to run locally instead of on Modal.