Fine-tuning

A pre-trained model is shown labelled examples, such as questions with ideal answers, and its parameters are updated to reproduce them. Full fine-tuning updates all parameters, while parameter-efficient methods such as LoRA (Hu et al., 2021) train small add-on weights, cutting cost. Instruction tuning and reinforcement learning from human feedback are forms of fine-tuning used to turn base models into helpful assistants.

Why it matters: fine-tuning can make outputs more consistent in tone and format, teach a narrow classification task or help a smaller, cheaper model approach a larger one on a specific job. For most business needs, though, better prompts and retrieval come first: they are cheaper, faster to change and keep information current, whereas knowledge learned in fine-tuning is frozen at training time.

Example: a company fine-tunes a model on thousands of past support replies so new drafts match its house style and ticket categories.

Limits: fine-tuning needs clean, representative data, evaluation and maintenance, and it is not a reliable way to add facts. Some vendors offer it only on certain models or plans.