Fine-tuning data preparation builds the pairs a later training job will consume: instruction and answer, or a short dialogue you are willing to imitate. I match the field names your trainer expects and keep the split from train, validation, and test. Fine-tuning here means the data, not the GPU run.

I do not invent a company voice from a handful of marketing lines. If the target behavior is not in your real answers, we either collect more or mark a synthetic slice under synthetic data. Preference pairs are in scope only when the brief has two ranked answers, not as a default. How to shape a supervised corpus is written up in SFT dataset design.

Acceptance is a file the named trainer can read, a sample of pairs with sources, and a validation slice that does not repeat train. Running the fine-tune and serving the model are AI implementation. A focused SFT set is often 3–6 weeks.