Supervised Fine-Tuning (SFT): Static Labels, Matching Completions

~12 min read

SFT fine-tunes a model on a fixed dataset of prompt-completion pairs, adjusting weights so the model's output matches the given completions — the book's baseline fine-tuning approach before RFT.

Supervised Fine-Tuning (SFT): Static Labels, Matching Completions is a Pro topic

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

Key points

  • SFT's process (per the book): start with a static labeled dataset of prompt-completion pairs, adjust weights to match those completions, deploy the best checkpoint
  • 'Static' means the correct answer for every example is fixed before training starts — someone (human or synthetic pipeline) already decided what's correct
  • Teacher-forcing trains the model against the known-correct completion at every step, rather than its own prior (possibly wrong) predictions
  • The book's key framing: SFT uses static data and often memorizes answers, rather than learning to discover new correct reasoning paths
  • SFT excels when 'correct' can be written down in advance as a specific target (support responses, code style, structured extraction)