What it means
Fine-tuning is the process of further training an existing model on a smaller, specialised dataset so that it behaves differently, for example using specific terminology, a particular writing style or a fixed output format. It changes the model itself, unlike techniques that only change the input.
Real-world example
A health insurer fine-tunes a model on 50,000 anonymised, correctly coded claim descriptions so that it suggests billing codes in the insurer's format more consistently. Before using it, the team checks that the training data was anonymised properly and compares the fine-tuned model against the original on a test set.
Why it matters for AI adoption
Fine-tuning can improve consistency for narrow, repetitive tasks, but it requires good training data, technical expertise and ongoing maintenance, and it does not reliably teach a model new facts that stay up to date. For questions about an organisation's own changing documents, retrieval is usually the better first option.
Often confused with
Retrieval-Augmented Generation (RAG)
RAG gives a model relevant information at the moment of answering, without changing the model, so updated documents are reflected immediately. Fine-tuning changes the model's behaviour through additional training and is better suited to style, format or specialised task behaviour than to keeping facts current.
Questions to ask
- Why is fine-tuning needed here, rather than better prompts or retrieval?
- What data would we use for fine-tuning, and are we allowed to use it for this purpose?