Specializing a model that already works
Fine-tuning takes a model that has already been trained and trains it further on additional data to suit a particular use. Instead of building from nothing, you take a capable general model and put it through specialist education.
Why bother
A general LLM handles most things competently but can struggle with an industry’s terminology, a company’s particular phrasing, or a required output format. Train it further on your own past support conversations and it can answer in your organization’s voice.
Compared with training from scratch, the data and compute required are dramatically smaller.
Call center automation that knows your product names and past interactions is usually the result of this kind of adjustment.
For terminology alone, including a glossary and examples in the prompt often solves the problem outright. Try instruction design first, and reach for additional training only when that is not enough — that is the usual order in practice.
How it differs from RAG
The other common way to give a model specialist knowledge is RAG, where the model retrieves documents and answers with them in view. Fine-tuning teaches knowledge and manner into the model itself; RAG lets it read the material each time. Choose RAG when the knowledge changes often, and fine-tuning when you want a particular style or format internalized.
Does it matter to individuals?
Fine-tuning itself is a developer technique, but the broader idea of customizing an assistant for a specific purpose has reached individuals through features such as GPTs. Recognizing the term as additional training on an existing model is enough.