SFT (Supervised Fine-Tuning)

The technical name for instruction tuning — supervised learning on (input, ideal-output) pairs.

Definition

SFT is the formal name for instruction tuning. The model is shown examples of correct outputs and trained to imitate them. SFT data is more expensive than preference data per item but produces sharper instruction-following behavior.

Examples

  • Writing the gold-standard answer to a hard physics problem
  • Producing model-quality summaries of long documents

Related terms

Looking for actual gigs that use SFT (Supervised Fine-Tuning)? Browse all live listings or filter by $100+/hr.