Supervised fine-tuning

A pre-trained model has learned broad patterns from vast data. Supervised fine-tuning (SFT) continues training that model on a curated set of input–output pairs — question/answer, instruction/response — so the weights tilt toward the desired behaviour. The hard part is the data: a small set of high-quality demonstrations outperforms a large noisy one, but "quality" is expensive to define and gather. SFT can also overfit — the model memorises the examples instead of generalising the intent. It is the first adaptation step between raw pre-training and a model that follows instructions.