What is fine-tuning?

Large AI models are trained on broad, general material. Fine-tuning takes one of these models and trains it further on a focused set of examples, usually pairs of inputs and ideal outputs, so its behavior shifts toward that task. The result is a customized version of the model that tends to follow a specific format, tone, or way of classifying things without long instructions each time.

For example, an insurer that receives a steady flow of claim descriptions might fine-tune a model on past claims labeled with their correct category. The fine-tuned model may then sort new claims into the right categories more consistently than a general model given only a written description of each category.

Fine-tuning is often confused with giving a model access to documents. It is mainly about behavior and style, not about adding facts that can be kept up to date. If the goal is for an AI tool to answer from the latest policies or product details, retrieval-augmented generation, which looks up relevant documents at the moment of each question, is usually the better fit, because documents can be updated without retraining.

Fine-tuning takes technical skill, good-quality example data, and ongoing upkeep, since the tuned model may need retraining when requirements change or a newer base model is adopted. For many everyday tasks, well-written prompts, examples placed inside the prompt, and connected documents achieve what is needed without fine-tuning at all.

An example

A company fine-tunes a model on its library of approved product descriptions so that new drafts match its house style.