AI glossary · Basic · Language models
SFT
SFT (supervised fine-tuning) is additional training that shows a pre-trained language model examples made of "instruction and model answer" pairs so that it learns to respond in the desired way. It is the most basic post-training step for making a chatbot answer questions in a conversational format.
About this term
- Level
- Basic · Technical terms you often meet in AI news
- Category
- Models & generation
- Fields
- Language models
- Added
- 2026-10-08
How it works
First, "instruction–correct answer" examples written or reviewed by people are collected. The model takes the instruction as input and predicts the correct answer one token at a time; the more it differs from the actual answer, the larger the error (loss) it receives, and it adjusts its parameters little by little. The principle is the same as pre-training, but the training material is "examples with a defined correct answer" rather than arbitrary text. By repeating this process, the model learns the format and tone for answering questions and the habit of following instructions.
Why it matters
A model that has only been pre-trained is good at continuing text, but when given a question it may list a string of similar questions instead of answering. SFT is the first step in turning such a model into an assistant that follows the user's instructions. Usually, preference learning such as RLHF or DPO, or reinforcement learning, is added after SFT to refine answer quality. It is also the first method used to give a model the tone and knowledge of a specific field such as law, medicine or customer support.
Things to keep in mind
Results depend heavily on data quality. If the examples are wrong or biased, the model learns them as they are. Because it works by imitating model answers, in situations not covered by the examples it can produce plausible but wrong answers, that is, hallucinations. Training too much on narrow data can also reduce the general abilities the model originally had, and to cut costs it is often used together with methods like LoRA that modify only some of the parameters.
Example
If a company organizes its support records into "customer question–agent model answer" form and applies SFT to a publicly available open-weight model, it can build a support chatbot that follows the company's tone and service rules. Research introduced on this site, such as 'T-Search: An open-weight multi-step search agent retriever based on Qwen3.6-35B-A3B,' also mentions SFT in the process of building a task-specific model from a public model.
Where it came up on this site
- Papers Kandinsky 6.0 Video: A foundation model that generates synchronized video and audio 2026-10-07
- Papers ALoDLM: A looped diffusion language model that adjusts iteration depth per token 2026-10-07
- briefing T-Search: An open-weight multi-step search agent retriever based on Qwen3.6-35B-A3B 2026-10-06