Skip to content
ML Academy

Introduced on day 43

Supervised fine-tuning (SFT)

Continuing to train a pretrained model on input-output example pairs so that it reproduces the demonstrated behaviour or format.

In detail

SFT adjusts weights with the same next-token objective, computed on the response part of curated examples. It is effective for consistent style, format and task behaviour and poor at adding large amounts of new, changing facts. Quality and consistency of the examples matter more than their number; it can also weaken safety behaviours, so safety tests are rerun after training.