Glossary

SFT

SFT, or supervised fine-tuning, trains a model on curated input-output examples so it learns a desired task, style, domain, or reasoning pattern. Good SFT data is explicit, consistent, and aligned to the behavior the model should show in production.

Request a Model Quality Audit Explore Corpus

What Makes SFT Data Useful

SFT data must show the model what good answers look like. It should include clear prompts, reliable answers, and enough domain variety to generalize.

Instruction-response pairs
Reasoning and explanation traces
Domain-specific answer formats

Where SFT Fits

SFT is commonly used after pre-training and before or alongside preference optimization, evaluation, and safety tuning.

Enterprise copilots
Reasoning models
Code and medical AI systems