What is Supervised Fine-Tuning?
SFT
It is the process of adapting a pre-trained artificial intelligence model to a specific task using labeled data pairs.
Overview
Supervised Fine-Tuning enables foundation models, which have acquired general language skills through large datasets, to be directed toward a specific area of expertise. In this process, direct sample data prepared in the form of questions and ideal responses are presented to the model. The model learns how to generate correct responses to given instructions by analyzing these labeled data pairs.
How it works
The model's weights are updated by small margins based on the input and output patterns in the prepared dataset. By calculating the error rate, the model's responses are made to approach the target responses. This process helps align the model with the desired format without experiencing overfitting.
Where it is used
It is used to ensure customer service bots respond appropriately to internal company documents and to make coding assistants produce output according to specific standards.
Commonly confused with
It can be confused with the general concept of Fine-Tuning. While Fine-Tuning is a general adaptation umbrella, Supervised Fine-Tuning represents the specific stage conducted with labeled instruction and response pairs.
Questions and answers
How much data is needed for Supervised Fine-Tuning?
Depending on the complexity of the task you are developing, generally between a few hundred and a few thousand quality data pairs are sufficient.
Does the SFT process make the model forget its general capabilities?
In misconfigured or overly prolonged training processes, the model may carry the risk of losing its general knowledge base.
Related terms
This explanation was written in plain language for TreScout with AI and machine-translated from the Turkish original · the Turkish version prevails. If something looks wrong or missing, write to hello@trescout.com. Read in Turkish →