When to Use SFT vs RFT: The Book's Decision Tree

~12 min read

The book's exact decision flowchart: labeled data or not? Verifiable outcome or not? Large or tiny dataset? Each branch points to a specific, concrete choice — SFT, RFT, or RLHF.

When to Use SFT vs RFT: The Book's Decision Tree is a Pro topic

Sign in, then upgrade to Pro or Power to unlock this topic and the full AI Engineering curriculum.

Key points

  • Root question: do you have labeled (ground-truth) data at all?
  • No labels -> is the task verifiable? Not verifiable -> RLHF (needs human preference signals); verifiable -> RFT (correctness auto-checked)
  • Have labels + large dataset -> SFT (direct imitation is efficient and stable with enough examples)
  • Have labels + tiny dataset -> does reasoning/CoT help the task? Yes -> RFT (explore via reward despite few labels); No -> SFT anyway
  • The tree converts a fuzzy 'which technique sounds better' choice into concrete questions about your own data and task