Glossary

AI Writing

Fine-Tuning

Further training a pre-trained model on additional task-specific examples to specialise its behaviour.

Definition

Fine-tuning takes a general-purpose pre-trained language model and continues its training on a smaller, focused dataset — typically pairs of prompts and the responses you want. The result is a variant of the model that is better at the specific style, format, or domain represented in the examples. Fine-tuning differs from prompting (no permanent change to the model) and from training from scratch (which requires vastly more data and compute).

Why it matters

Fine-tuning was the first standard way to specialise a language model. It still has a role — for narrow tasks with stable inputs and outputs (classification, structured extraction, consistent style) — but for many use cases it has been displaced by better prompting and retrieval-augmented generation, which are cheaper, faster to iterate, and easier to govern. For writing teams specifically, fine-tuning a model on a brand-voice corpus can produce consistent on-brand drafts at scale. The trade-off is operational: the fine-tuned model must be maintained, retrained as voice evolves, and re-tested whenever the base model is updated.

Examples

Good fit

Fine-tuning a small open-source model to extract structured product attributes from messy supplier feeds.

Often overkill

Fine-tuning a frontier model to write 'in our brand voice' when a strong system prompt plus three brand examples in context achieves nearly the same result.

Frequently asked

How much data do I need to fine-tune?

For style and format tasks, often a few hundred to a few thousand examples suffice. For factual capability, far more is needed and prompting/RAG is usually a better lever.

Is fine-tuning permanent?

The resulting model is a separate artifact. The original base model is unchanged.

Can fine-tuning leak training data?

Yes, if the dataset contains sensitive content. Treat fine-tuning corpora with the same care as any database of regulated data.

Related terms

Put the concept to work

Open the rewrite engine and apply this principle to a draft of your own.

Try a rewrite