# What is Fine-tuning?

Also called model adaptation.

Fine-tuning continues training an already-trained model on a smaller, targeted dataset so it performs better on a specific task, domain, or format. It changes the model's parameters, which distinguishes it from prompting. Typical uses include enforcing a house style, matching a rigid output schema, or teaching specialized terminology that appears rarely in general training data.

Full fine-tuning updates every parameter and requires memory comparable to training the model in the first place. Parameter-efficient methods, of which low-rank adaptation is the best known, instead freeze the original weights and train a small set of additional ones. These adapters are a tiny fraction of the size, cost far less to train, and can be swapped per customer or per task.

Fine-tuning reliably teaches behavior and format. It is much less reliable for teaching facts, because new information competes with everything absorbed during pre-training and may be recalled inconsistently or blended with older material. For knowledge that changes over time or must be cited to a source, retrieving documents into the prompt is usually the better approach.

Data quality dominates outcomes far more than data volume. A few hundred carefully curated, mutually consistent examples generally outperform tens of thousands of noisy ones, and contradictory examples teach contradiction. Fine-tuning can also degrade abilities outside the target task, an effect known as catastrophic forgetting, so evaluating on general tasks after training is a necessary step.

Cost extends well past the training run itself. A fine-tuned model must be hosted, evaluated, and recreated whenever the underlying base is upgraded, which happens often. Teams commonly start with careful prompting and retrieval, measure the result, and fine-tune only when a specific measured gap remains that prompting has demonstrably failed to close.

## Key points

- Changes model parameters, unlike prompting, which changes only input.
- Adapter methods train a small add-on instead of all weights.
- Best for behavior, tone, and format; weak for volatile facts.
- A few hundred clean examples beat thousands of noisy ones.
- Fine-tuned models must be rebuilt when the base model changes.

## In practice

A support team needs every reply to open with an apology, cite a policy number, and end with a ticket reference. Prompting gets this right most of the time but drifts on long conversations. The team collects eight hundred approved replies and fine-tunes a smaller model on them. The format becomes consistent, the prompt gets shorter, and each request consumes fewer tokens.

## Related terms

- [Pre-training](/en/glossary/pre-training)
- [Instruction Tuning](/en/glossary/instruction-tuning)
- [Prompt Engineering](/en/glossary/prompt-engineering)
- [Small Language Model](/en/glossary/small-language-model)
- [Reinforcement Learning from Human Feedback](/en/glossary/reinforcement-learning-from-human-feedback)

[Back to the AI Glossary](/en/glossary)