GuideAdvanced

Fine-Tuning and Distillation for LLMs

Fine-tuning a model means changing its weights with your own examples so it does a task in a fixed way without needing a long prompt. Sometimes it is the best decision in the project, and sometimes it is a cost nobody will maintain. This guide teaches you to tell which one is yours, with numbers, and to do it well when it pays: decide against prompting and RAG, prepare the data, evaluate before and after, understand LoRA and QLoRA from the inside, train open weights in the cloud or on a Mac, use providers' fine-tuning knowing what you risk, go beyond SFT with preferences and reinforcement, and distill a large model into a small one. You learn it with **Halyard Energy**, a fictional electric utility that classifies 1.5 million customer messages a month into a JSON label with an intent, a priority and three extracted fields. Today a large model does it with a 653-token prompt. You will build everything that decides whether a small fine-tuned model can replace it: a clean, leak-free dataset from a dirty export, a golden set with audited labels, the evaluation with slices, rare classes, intervals and regressions outside the task, the parameter and memory math of a real model from its `config.json`, the training configurations for TRL + PEFT and for mlx-lm with their preflight, the providers' requests validated without sending them, the preference pairs, a grader that cannot be gamed, and the dataset labeled by the teacher model. The lab runs on any laptop, with no GPU, no model downloads and no API keys. LoRA and distillation are trained for real on numpy toys that finish in seconds; the model answers being evaluated are simulated with fixed error profiles and labeled that way in every lesson, and what is real is the code that evaluates them. Every provider fact, price and version is verified in the official documentation as of October 2026, including what changed the most: several providers have closed or are closing their fine-tuning, and the guide tells you which.