Fine-Tuning

Fine-tuning is the process of taking an existing model and continuing to train it on a curated set of examples so it learns to behave better for a specific task, domain, style, or output format. Unlike prompting — which steers model behavior through instructions — fine-tuning changes the model's parameters directly. The result is a model that reliably follows patterns that are difficult to enforce through prompting alone: specialized classification taxonomies, regulated communication styles, or structured output formats that general models handle inconsistently.

Fine-tuning is frequently proposed as the solution to customization problems that retrieval or better prompting would solve more simply. It adds governance overhead — training data curation, evaluation, versioning, retraining cycles — and it doesn't give the model updated knowledge or guarantee factual accuracy. Poor training examples can make a model consistently wrong in a very polished way. Before approving a fine-tuning project, the key question is whether the customization problem is actually a behavior problem (where fine-tuning helps) or a knowledge problem (where retrieval is the right answer).