What is fine-tuning?

Fine-tuning is taking a model that already exists and training it a bit further on examples you supply. You are not building a model from scratch — that costs millions. You are adjusting one that already speaks English to speak it the way you want.

What it is actually good at

Fine-tuning changes behaviour, reliably:

Format. Always return JSON in this exact shape. Always use this structure.
Tone. Sound like our support team, not like a press release.
A narrow task done consistently. Classify tickets into these nine categories.
Dropping the preamble. Stop saying "Certainly! Here's a…" every time.

The pattern: things you could explain in a prompt but would rather not repeat on every request, and which need to be consistent rather than approximately right.

What it is bad at

Teaching facts. This is the expensive mistake, and it's made constantly.

Fine-tuning on your product documentation does not install that documentation in the model. It teaches the model to produce text shaped like your documentation — the register, the vocabulary, the structure. Asked a specific question, it will now generate confident, on-brand, plausibly-formatted answers that may be entirely invented. You have made hallucinations harder to spot, not less frequent.

If the answer lives in a document, retrieve the document. That's RAG.

Choosing between the three options

Your problem	Reach for
It gets the facts wrong	RAG
Facts change often	RAG
You need to cite sources	RAG
The format is inconsistent	Fine-tuning
The tone is wrong	Fine-tuning
Your prompt is enormous and repetitive	Fine-tuning
You have fewer than ~100 good examples	Prompting
You aren't sure	Prompting, then RAG, then fine-tuning

Work down that list in order. Prompting costs nothing to try. Fine-tuning costs money, a dataset, and a retraining step every time you change your mind.

LoRA, and why fine-tuning got cheap

Full fine-tuning updates every weight in the model — expensive, and it produces a whole new copy. The 2021 LoRA technique took a different route: freeze the original model and train a small set of additional weights alongside it. The result is a small adapter file rather than a new model.

This is most of why fine-tuning became practical for ordinary teams. It's also why you can host one base model and swap between many fine-tunes cheaply.

Before you fine-tune

Your dataset is the whole job. A few hundred genuinely good examples beat tens of thousands of mediocre ones, and the model will faithfully learn any inconsistency you leave in. Most failed fine-tuning projects are dataset problems, discovered late.

What we addedA decision table for fine-tuning vs RAG vs prompting, and the specific reason fine-tuning fails at teaching facts.

This article was researched and drafted with AI assistance from the sources listed below, then checked and edited by Fiqhro Dedhen before publication. How we work.