Fine-Tune or Just Prompt?

Fine-Tuning

Part of: Fine-Tuning & Evals

Someone on your team says "the model keeps getting the tone wrong, let's fine-tune it." Nine times out of ten, that's the wrong first move. You probably have a prompt problem, not a model problem, and you're about to spend two weeks and a dataset budget to fix something a paragraph of instructions would have solved in an afternoon. What it is Fine-tuning takes a base model and nudges its weights using your own examples. You hand it hundreds or thousands of input→output pairs, and it shifts to favor the patterns in your data. The model doesn't gain new knowledge so much as new habits , a house style, a rigid output format, a domain vocabulary, a way of refusing. Contrast that with prompting , which changes nothing about the model. You just write better instructions at request time. Same weights, different message. And few-shot prompting sits in between: you drop a handful of example input/output pairs straight into the prompt to demonstrate the behavior, no training required. The mistake teams make is treating fine-tuning as the default fix. It's the last resort. How it works The honest decision rule is a ladder, climb it in order, stop at the first rung that works: 1. Plain prompt.

Challenge: The Fine-Tune Router