Fine-tuning is the process of taking a pre-trained model and training it a little more on a smaller, task-specific dataset so it behaves better for your use case.
Pre-trained models are general-purpose: they know a lot, but not your company’s style, domain vocabulary, or output format. Fine-tuning helps when you want the model to:
In practice, teams often try prompting first, then retrieval, and reach for fine-tuning when the behavior needs to be more stable or the task is repeated enough to justify training.
The basic idea is simple: start with a model that already learned broad patterns from large-scale training, then continue training it on examples that reflect the behavior you want.
Usually, that means providing input-output pairs, such as user requests and ideal responses. The model updates its weights slightly so it becomes more likely to produce those target outputs in similar situations.
There are different kinds of fine-tuning. In common language, people often mean supervised fine-tuning, where examples are labeled and the model learns from them directly. In research and product practice, you may also see parameter-efficient methods like LoRA, which adapt only part of the model rather than all of it. The exact method matters, but the core idea is the same: specialize a general model.
Fine-tuning does not give the model new facts in a reliable, always-up-to-date way. It changes behavior more than it changes knowledge. If you need current or factual information, retrieval or tools are often a better fit.
Suppose you want a support chatbot that always answers in this format:
Answer:
1. Short response
2. Next step
3. Escalate if needed
You collect a few hundred examples of good support replies in that style and fine-tune the model on them. After training, the model is more likely to produce that same structure without needing a long prompt every time.