SFT

SFT takes a pre-trained language model and continues training it on curated example conversations so it learns a specific style, format, or domain, usually as the step before preference tuning like RLHF. For organizations, fine-tuning on internal data creates a new asset to protect: the resulting model can memorize and regurgitate fragments of its training data, so the tuned weights deserve the same sensitivity classification as the data that shaped them.