Automation and AI

Fine-Tuning

Also called Model tuning, custom model

Extra training of a finished model on your own examples so it adopts a format, tone or task by default.

Quick facts: Fine-Tuning

Category
Automation and AI
Also called
Model tuning, custom model
Level
Advanced
Affects
Output consistency, running cost, model size you can get away with
Where to see it
Model provider consoles: OpenAI platform, Google Vertex AI, Azure AI Foundry
In this article4
  1. How fine-tuning works
  2. Why fine-tuning matters
  3. Where fine-tuning goes wrong
  4. How to act on it

How fine-tuning works

A general model has already been trained on a broad body of text before you ever meet it. Fine-tuning takes that finished model and continues training it on a much smaller, curated set of examples that demonstrate the behaviour you want: your past support replies, your product descriptions written the way you write them, pairs of rough notes and the tidy summary you would have made of them. The model’s internal weights shift a little towards those examples.

What you end up with is a private variant that leans your way without being told each time. It has learned a pattern, a format and a register. It has not reliably learned new facts, and that distinction is where most of the disappointment starts.

Why fine-tuning matters

For marketing work it is mainly a cost and consistency decision. If a longer, clearer prompt gets you an acceptable answer, use the prompt: it is cheaper, you can change it in a minute, and it needs no engineering. Fine-tuning earns its place when the same instruction is repeated at volume, when the prompt has grown so long that it is expensive on every single call, or when a house style is too subtle to describe in words but easy to show with examples.

It is also a way to get a small, cheap model behaving well enough for one narrow job, rather than paying a large model to do something simple.

Where fine-tuning goes wrong

The commonest mistake is reaching for it to solve a knowledge problem. If the model needs to know your current prices, stock or refund policy, that is a job for retrieval-augmented generation — supplying the facts at the moment of the request — not for training. Facts pressed into weights go stale and cannot be corrected without training again.

The second is training on a messy set. A tuned model copies whatever it is shown, including the sloppy replies, the old brand name and the message that upset a customer. A small, clean, consistent set beats a big scruffy one every time.

The third is forgetting maintenance. A tuned model is frozen on the day you tuned it, while your offer, your tone and the underlying base model all keep moving.

How to act on it

Start without it. Write the prompt properly, put a few worked examples inside the prompt, and connect the model to your live data. Only when that ceiling is genuinely reached should you gather training examples deliberately — approved outputs, not raw archives — and keep the set in version control so you can retrain from something known to be good.

Decide before you start who reviews what comes out, because a tuned model produces confident work in your own voice, which is exactly the sort that gets published without a second look. If you are weighing this against simpler options, an honest review of your AI automation setup usually turns up cheaper wins first.

Do and do not

Do

  • Fix the prompt and the data source first
  • Train on approved outputs, never raw archives
  • Keep the training set in version control

Do not

  • Fine-tune to teach a model current facts
  • Assume a tuned model stays correct forever
  • Put customer personal data in a training file

Questions people ask about this

Is fine-tuning how I teach a model about my products?

Usually no. Fine-tuning teaches a model a pattern, a format and a tone rather than dependable facts, and anything baked into the weights goes stale the moment your prices or policies change. To give a model current information about your business, connect it to that information at the moment of the request instead.

How many examples do I need to fine-tune a model?

There is no fixed answer, and quality decides it more than quantity does. A small set of clean, consistent, approved outputs teaches a clearer pattern than a large archive full of contradictions, outdated wording and replies you would not send today. Start with the best examples you have, test the result, and add more only where the output is still wrong.

Is a fine-tuned model private to my business?

The tuned variant is yours, but the terms depend entirely on the provider, so read them before you upload anything. The bigger practical question is what goes into the training file. Customer names, phone numbers and enquiry details do not belong there unless you have a lawful basis for using them and have stripped out everything you do not need.

Related terms

Found this useful?

Share it, or ask an AI to summarise it

Back to the glossary

Knowing the term is the easy part

Applying it to your own site and budget is the work. Book a call and I will tell you what actually applies to you.