What Is Fine-Tuning?

Fine-tuning continues training an existing model on examples of the behaviour you want, adjusting its weights. It is how you teach a model to respond in a particular structure, tone or classification. It is not a reliable way to teach it facts.

1 min read Updated August 29, 2026

What fine-tuning teaches well

Output structure, where the same shape is needed every time and prompting gets close but not close enough.
Tone and register, across thousands of generations, where describing the style in a prompt would cost more than teaching it once.
Classification, where a small fine-tuned model regularly beats a large prompted one on both cost and latency.
Domain phrasing, so the model uses your vocabulary rather than a generic equivalent.

What fine-tuning teaches badly

Facts. Training adjusts the likelihood of a phrasing rather than storing something you can look up, so a fine-tuned model will often produce a confident answer that resembles the truth, with no source to check it against. Wrong, fluent and untraceable is the worst combination a production system can have. Facts belong in documents, retrieved at the moment of the question.

What fine-tuning actually costs

The compute is rarely the expensive part. Preparing training data is: gathering examples, making them consistent, and holding back a set to evaluate against. That work has to be repeated whenever the desired behaviour changes, which is the cost people underestimate when they choose it over retrieval.

When is fine-tuning worth it?

When the behaviour is stable, the volume is high, and prompting has already been pushed as far as it goes. If your prompt is longer than the input it processes, that instruction probably belongs in the weights. If the thing you want the model to know changes weekly, it does not.

RAG vs fine-tuning: the full comparison, including when to use both.
What is RAG?: the approach to try first.