All terms
SFT (Supervised Fine-Tuning)
Training a model directly on example input-output pairs written or corrected by humans, rather than on ranked preferences.
SFT is usually the first tuning step after a model's initial large-scale training: humans write or edit "ideal" responses to prompts, and the model is trained to imitate them directly. It's the more literal sibling of RLHF - instead of ranking multiple outputs, you're producing (or fixing) the single example the model should learn from.
Example
You're given a customer email and asked to write the exact reply the model should learn to produce, rather than to rank several AI-drafted replies.