Supervised Fine-Tuning

What is supervised fine-tuning?

A base model has read oceans of text and can "continue" a sentence, but it can't yet "follow orders" — ask it a question and it might just keep writing instead of answering. Supervised fine-tuning (SFT) trains it for another round on carefully prepared question–answer pairs, so it learns to answer what's asked, and answer well.

How is it different from pre-training?

Pre-training: learning language itself
At this stage the model is just reading at scale, learning "what usually comes next". That's grammar and common sense — but not yet a service mindset.
SFT: learning how to respond
Using Q&A and instruction samples, SFT teaches the model that "this user needs help, so give a direct, useful answer". Only after this does the model start to feel like an assistant.

Where does the data come from?

SFT quality depends heavily on data. Good instruction data often comes from human writing, curated high-quality answers, or even stronger models generating drafts that humans then verify. Cleaner and more diverse data means a more reliable model.

What it does — and its limits

SFT quickly teaches format and instruction-following at lower cost than RLHF, but it only "imitates the model answer" — it can't yet tell "good" from "better". That's where RLHF steps in.

Bottom line: supervised fine-tuning shows an AI the "model homework" and has it learn how to answer accordingly.