AI glossary for businesses
Fine-tuning
In short
Fine-tuning means taking an AI model already trained on general data and training it further on a smaller set of your own examples. The goal is to make it follow a specific style, format or task more consistently. It changes how the model behaves, but it is not the best way to teach it facts that change often.
How fine-tuning works
Imagine hiring an experienced chef. They already know how to cook. You do not send them back to culinary school. You show them your restaurant's house recipes and plating, and after some practice they produce your dishes the way you want. Fine-tuning works the same way: the model keeps its general skills and learns your specific patterns.
In practice, you prepare a set of examples, often a few hundred or more, each with an input and the ideal output. The provider runs extra training on them and gives you a custom version of the model. OpenAI, Google and other providers offer fine-tuning for some of their models, and open-source models such as Llama or Mistral can be fine-tuned on your own infrastructure.
In practice for a small business
An online furniture shop has 5,000 product descriptions that the team has approved over the years. They want new descriptions to follow exactly the same structure, length and tone. Fine-tuning a model on those approved pairs (product data in, finished description out) can make the output more consistent than a prompt alone.
But for most SMEs, fine-tuning is not the first step. Try these in order:
a clear prompt with three or four good examples;
RAG, if the model needs your current prices, policies or stock data;
fine-tuning, only if the first two still give inconsistent results at high volume.
The article build, buy or customise AI helps with this decision.
Why it matters and what to watch out for
Data quality: the model copies your examples, including their mistakes. Clean data matters more than lots of data.
Cost and effort: you pay for training and often for hosting or using the custom model. Someone must also prepare data and test results.
Maintenance: when the provider releases a better base model, you may need to fine-tune again.
Personal data: once personal data is used in training, it is hard to remove. Strip names and contact details from examples. If personal data is involved at scale, a DPIA under GDPR may be needed; check with your DPO.
Facts go stale: a fine-tuned model does not update itself when your prices change.