Fine-tuning

Fine-tuning means continuing the training of a model that has already learned the basics. A large language model is first pretrained on a vast, general corpus until it has broad command of language and reasoning. Fine-tuning then takes that base model and trains it a little more on a curated set of examples that show the specific behavior you want — adjusting the model's weights so it leans toward your task, domain, format, or voice.

How it works: you assemble training pairs that demonstrate the target behavior — say, customer messages mapped to the exact JSON your system expects, or prompts paired with answers in your brand's tone. During fine-tuning, the model's parameters shift to reduce the gap between its predictions and your examples. Lightweight methods such as LoRA update only a small slice of the weights, which keeps the process cheaper and faster. Afterward, every request at inference time uses the adjusted weights, so the new behavior is baked in rather than re-explained in each prompt.

Why it matters: fine-tuning earns its keep when you need consistent behavior that instructions alone can't guarantee — a rigid output schema, a specialized classification skill, or a distinctive style held across thousands of calls. It is less suited to supplying fresh facts, since knowledge baked into weights goes stale and is expensive to refresh; for that, teams reach for retrieval-augmented generation (RAG) instead, and often combine the two — fine-tune for behavior, retrieve for knowledge.

Concrete example: a legal team wants summaries that always follow a fixed five-section template. Prompting gets close but drifts. By fine-tuning on a few hundred hand-formatted summaries, the model learns the template so well that it reproduces the structure reliably from a short instruction — saving prompt tokens and removing the formatting errors that crept in before.