Fine-tuning
Also known as: Model fine-tuning · Retraining
Fine-tuning means further training an already-trained model on your own data to adjust its behavior or style.
A language model is first trained generally on broad amounts of text. Fine-tuning then takes that finished model and continues training it on a smaller, targeted dataset, so it answers in a particular format, adopts a particular tone, or uses an industry's terminology more consistently. The goal is adapting to a narrow use case, not teaching entirely new world knowledge.
In practice, fine-tuning needs a carefully assembled training dataset, computing power for the training run, and an evaluation afterward to check whether the behavior actually changed as intended. The resulting model also needs upkeep: when the provider updates the base model, fine-tuning often has to be redone.
The limiting factor is that fine-tuning is not a reliable way to teach a model new facts, and with too little or one-sided training data, the model's general capability outside the trained area often suffers. It is also more effort and cost to operate and maintain than many people expect.
What it means in practice
For most businesses of around fifty people, fine-tuning is honestly rarely the right path: anyone who needs current, checked answers from their own documents usually gets there faster, cheaper and with less maintenance using RAG. Fine-tuning only pays off for a very specific, recurring format that RAG alone doesn't cover.
Is this handled properly at your site?
We look at how it actually stands with you — and say honestly whether anything needs doing.
Related terms
All termsA term from your quote missing here?
Send us the passage you do not follow. We will explain it — with no obligation to order anything.