Supervised Fine-Tuning (SFT)

Full definition

After pretraining on raw text, base models are fluent but not helpful. SFT teaches the model the response format users want — answer questions, follow instructions, refuse unsafe requests. The dataset is usually 10k-1M high-quality demonstrations. SFT is the first of three post-training stages in the InstructGPT/ChatGPT recipe: SFT, then reward modeling, then RLHF. SFT alone can recover 80%+ of the gains from full RLHF on many tasks at much lower cost.

Why it matters

SFT is the cheapest, highest-ROI customization step for enterprises. Most "fine-tuned for our domain" claims are SFT, not RLHF. Buyers should ask vendors how many SFT examples were used and how they were curated — quality dominates quantity above ~10k examples.

Example

A pharma company runs SFT on Llama 3 70B with 25k pairs of (clinical question, regulator-approved answer); the resulting model passes internal compliance review at 3x the rate of the base model.

Related terms

• Reinforcement Learning from Human Feedback (RLHF)
  Reinforcement Learning from Human Feedback (RLHF) is a training technique that aligns a language model's outputs with human preferences by training a reward model on human comparisons of candidate responses, then optimizing the LLM against that reward.
• Fine-tuning
  Fine-tuning is the process of further training a pretrained model on a smaller, task-specific dataset so it specializes in a particular style, domain, or behavior — without retraining from scratch.
• Pretraining
  Pretraining is the first, most compute-intensive stage of training a foundation model — typically self-supervised next-token prediction over trillions of tokens of text, code, and other modalities — that produces the base model later fine-tuned for specific tasks.
• LoRA (Low-Rank Adaptation)
  LoRA (Low-Rank Adaptation) is a parameter-efficient fine-tuning method that freezes a pretrained model's weights and injects small trainable low-rank matrices into selected layers — reducing fine-tuning cost by 100-1000x with negligible quality loss on most tasks.