The model learns the exact string shape you give it. Message roles, field order, and punctuation are all part of what it copies, so a mismatch between how you format the training set and how you format the live request is a bug that shows up as garbage output.
About 1 min of reading
Share one template.
The builder
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.
Write the function that builds a request once and use it for training and for serving. Two copies of a template is where the drift lives, and drift is the failure.
Parse every expected output before training. A set with unparseable outputs trains on noise, and the run looks successful while the model learns nothing usable.
If serving sends a system message plus one user message, the training examples must have that same shape. Train on a bare string and serve with roles, and the model sees a distribution it never learned.
If you want JSON, the expected output is valid JSON with the exact keys. A set where half the outputs have an extra key teaches a model that sometimes emits that key.
Print three examples in the exact bytes that will be sent. That check is cheap and it catches the mismatch that a training run then wastes hours to reveal.
Each one is defined here on this site. Open the entry for the full definition and the other words it points at.