Fine-tuning is supervised: it needs pairs of an input and the output you want. That set is the specification of the job, written as examples instead of as rules, and its quality is what the whole run is made of.
About 1 min of reading
Write three pairs.
A pair
Before you run it, say out loud what it will print. Then open the answer.
A refusal pair
Before you run it, say out loud what it will print. Then open the answer.
Read the first example one line at a time. A line that prints shows what appears on screen; the rest set things up quietly.
Include inputs the model should decline, with the decline as the expected output. A set with no refusals teaches the model to answer everything, including questions it has no basis for.
Hold back examples the model never trains on and evaluate on those. Measuring on the training set tells you the model memorised, which is a different and worse outcome.
Each example has the input as a real user would send it and the output exactly as you want it. A set of documents is not a training set, and feeding one produces a model that rambles about them.
Examples gathered from real traffic carry the mess: typos, half questions, extra context. A hand written set is clean, and the model then fails on the real thing.
The set must include the short, the long, the unusual, and the ones that should be refused. A set of fifty near copies of one question teaches one case and nothing else.
Each one is defined here on this site. Open the entry for the full definition and the other words it points at.