Fine-tuning
Also called fine tuning, model fine-tuning, supervised fine-tuning
Training an existing AI model further on your own examples to change how it behaves.
Fine-tuning takes a pre-trained model and trains it further on a smaller, focused dataset, adjusting its parameters so it gets better at a particular task or style. Many providers offer fine-tuning through their APIs, and open-weights models can be fine-tuned on your own hardware. Techniques like LoRA make it cheaper by training only a small number of extra parameters.
Fine-tuning is good at teaching form: a consistent output format, a house style, a specialised classification task, or shorter prompts for a task you run millions of times. It can also make a smaller, cheaper model perform like a larger one on a narrow task.
It is less good at teaching facts. A model fine-tuned on your product docs will not reliably recall details from them and may confidently mix them up. For knowledge that changes or must be accurate, retrieval (RAG) is usually the better choice. Most teams should try careful prompting and few-shot examples first; fine-tuning is worth it when those reach their limit and you have hundreds or thousands of high-quality examples.
Example
A company classifies 2 million support emails a month into 40 categories. A small fine-tuned model, trained on 8,000 labelled emails, matches a large model's accuracy at a fraction of the cost and with much shorter prompts.
Common mistake
Fine-tuning to fix a problem that a better prompt would solve. Fine-tuning needs data collection, training runs and evaluation, and must often be redone when the base model changes. Exhaust prompting and examples first, and measure both against the same test set.