SFT Supervised Fine-Tuning
The first finetuning step after pre-training — train on prompt+response pairs so the model learns to follow instructions.
Before RLHF/DPO. The dataset is hand-written or curated examples: "question X → good answer Y". Smaller dataset (10k-100k examples) but high quality. Step 1 → SFT, step 2 → RLHF or DPO. Together = an "Instruct" model.