Definition
Supervised fine-tuning, commonly shortened to SFT, continues training a pretrained model on examples that pair an input with a desired output. Training adjusts model parameters to increase the likelihood of the target response given the input.
The examples may represent instructions and answers, conversations, classifications, structured transformations, or other task-specific behavior. SFT describes the supervised learning setup and objective, not how many parameters must be updated.
Simple example
A team prepares reviewed examples containing a user support request and an approved response with the correct tone, escalation decision, and output fields. During SFT, the model sees the request as input and the approved response as the target sequence. Evaluation then checks the adapted model on separate requests that were not used for training.
The same SFT dataset could be used for full-parameter fine-tuning or with a parameter-efficient method such as LoRA.
Why it matters
SFT gives teams a direct way to demonstrate desired behavior rather than expressing every pattern through prompt instructions. It is a common step when adapting a general model to a task or preparing it for later preference optimization.
Its effectiveness depends heavily on example quality, coverage, consistency, and separation between training and evaluation data.
One important nuance
SFT trains a model to reproduce target responses, including errors and inconsistencies. A low training loss does not prove that the model generalizes or remains reliable outside the example distribution. Check training data for unintended duplicates and conflicting examples. Protect sensitive information, keep evaluation cases separate, and test important capabilities for regressions. SFT is distinct from LoRA: SFT specifies the learning signal, while LoRA specifies a parameter-efficient way to apply updates.