What the data actually looks like
A fine-tuning dataset for an instruction-following model is typically a collection of examples, each pairing an input, an instruction or a user message, with the exact output you want the model to produce in response. Every example is a small demonstration of the behavior you're trying to teach, and the model learns the pattern across all of them, not by memorizing any single one, but by picking up on what's consistent across the whole set.
This means the format and structure of your examples matters as much as their content. If you want the model to always answer in three bullet points, every training example needs to actually be formatted that way. If half your examples use bullet points and half use paragraphs, the model has no consistent pattern to learn, and its output after training will be just as inconsistent as before.