Training

Supervised Fine-Tuning (SFT)

Supervised fine-tuning (SFT) continues training a pretrained model on high-quality labeled examples of desired behavior. After pretraining, models are often poor at following instructions. SFT on instruction-following examples dramatically improves this. SFT is cheaper than pretraining but still requires curated data. High-quality SFT data has outsized importance. SFT can overfit to specific formats or styles if not careful. Most deployed models use SFT before RLHF. SFT is the practical bridge between pretrained models and useful assistants.