Fine-tuning
Fine-tuning is the process of taking a model that was already trained on a broad task and training it further on a narrower, labelled set of examples, so it adapts its behaviour — tone, format or task-following — to a specific use without learning new facts from scratch.
Submit a project Compare fine-tuning against retrievalReviewed by Huy Nguyen (David), CEO · Updated 15 Oct 2026 · 3 min read
This term is one of the entries in the OutsourcingVN glossary of words buyers meet while deciding how an AI feature should use their data.
Why does fine-tuning matter?
Fine-tuning is the technique that lets a general-purpose model start acting like a specialist without starting from zero: a team trains a model that already understands language in general further on examples of the specific behaviour it wants, rather than building and training a new model from nothing. OutsourcingVN is operated by Netbase JSC, and a buyer who confuses fine-tuning with teaching the model new facts tends to pick the more expensive route for a problem a cheaper lookup would have solved.
What is an example of fine-tuning in practice?
A support team wants every reply to follow a strict three-part format: acknowledge, resolve, close. Prompting alone drifts from the format on about one reply in five. The team collects three hundred past replies that already follow the format correctly, labels them as examples, and fine-tunes the base model on that set. The fine-tuned model now follows the format far more consistently, but it still does not know this week's product changes — those still have to come from a lookup at answer time, not from the fine-tuning run.
What are the common misreadings of fine-tuning?
Fine-tuning is not a way to keep a model's knowledge current: once training ends, the model's facts freeze at that point, and a changing fact still needs a lookup at answer time rather than a retrain. It is also not free of the data problem — a fine-tuned model is only as consistent as the examples it was shown, and a handful of inconsistent ones will show up as inconsistent behaviour later. And fine-tuning is not the same decision as choosing which base model to use: the base model choice happens first, and fine-tuning is a further step applied to whichever base model was chosen.
Plan the next step for your project
Common questions
No fixed number applies everywhere, but a task usually needs from the low hundreds to a few thousand labelled examples before a fine-tuned model reliably clears an evaluation set.
Yes. A previous model version should stay available, so a fine-tuning run that performs worse can be rolled back without losing the base model.
Which decisions does fine-tuning affect?
Choosing fine-tuning over retrieval, or combining the two, is the subject of the RAG vs fine-tuning guide, which compares the levers by knowledge freshness, behaviour change, data volume, evaluation and running cost; the emerging technology adoption guide covers the same bounded-experiment method for fine-tuning alongside any other new AI capability. When a reply's facts or its format are not behaving the way your product needs, describe the behaviour you want changed, and Netbase will respond through Netbase's own outsourcing-services platform; submit a project when you are ready.