Supervised Finetuning

Supervised finetuning (SFT) is the most straightforward way to teach a model new behavior: you show it examples of the right answer, and it learns to replicate that pattern. For instance, you might create hundreds of examples where a user asks a question and a human expert writes the ideal response — then train the model on those pairs. SFT is often the first step before RLHF, giving the model a solid behavioral baseline before preference-based refinement begins. The quality of the examples matters enormously; a model trained on mediocre demonstrations will produce mediocre outputs, no matter how many examples you provide.