What is fine-tuning?

Fine-tuning is further training an existing AI model on a set of examples so that it behaves differently from then on. The examples, usually hundreds to thousands, show an input and the output wanted, such as a house writing style or a strict format. A prompt affects one request. Fine-tuning changes how the model responds every time.

Why it matters for a business

Fine-tuning changes behaviour, not knowledge. Using it to teach a model a company’s facts is a recurring, expensive mistake: the model picks up the tone of the documents and still invents the details. Facts belong in retrieval (RAG), which can be updated the same day a document changes. Fine-tuning earns its cost on a high-volume task where a long prompt on every call would be wasteful, after prompting and retrieval have genuinely run out of room. A fine-tune is also tied to one base model, so it has to be redone when a better base ships.

Example

Let’s say a firm fine-tunes a model on 5 years of its project documents, expecting it to know the firm’s clients. What comes back writes in the house style and still invents client details. The gap was knowledge, so the answer was retrieval.

Go deeper

RAG vs. fine-tuning vs. prompting, and what 'train it on our data' means

Related terms

• RAG (retrieval-augmented generation)
• Prompt engineering
• RLHF (reinforcement learning from human feedback)
• Foundation model
• Open-weight model

All terms, A to Z