supervised fine-tuning (SFT)

advancedwidespread

Further training of an already trained language model on question-answer pairs to perform a specific task or style more reliably.

In plain terms

Imagine a newly graduated, well-educated colleague joining your company. They know many things but do not know your internal rules. You give them a few hundred past cases with the correct solutions and say: this is how we do it here. Supervised fine-tuning is the same with a language model.

The term "supervised" means that every practice exercise comes with a pre-defined, accepted sample answer. This is written by a human or generated by another model, and ideally reviewed by a human. The model does not guess on its own; it is shown the correct output and aligns with it.

It is useful for turning a general model into a specialized one: to maintain the company's tone, always respond in the same format, or better understand a narrow field of expertise. This turns a raw text-completion model into a question-answering assistant.

It differs from a good prompt in that the model itself changes permanently here, not just what you ask of it. Therefore, preparation is more expensive and slower, but afterward you do not need to explain in detail every time what you expect.

How it works

As a starting point, you take a pre-trained model that has learned the language and much general knowledge from a huge amount of text.
You compile a set of examples: inputs (a question, an instruction, a document) and the expected answers that go with them, typically from a few hundred to tens of thousands of pairs.
The model receives the input and tries to predict the next text segment of the sample answer step by step. The system measures how much the prediction deviated from the sample answer.
Based on the deviation, the model's internal settings (weights) are adjusted in small steps to get closer to the pattern next time. This is repeated many times over the entire training set.
At the end, the model is tested on examples it did not see during training to determine whether it truly learned the task or simply memorised the patterns.

An everyday example

An insurer's customer service chatbot addresses you informally, answers briefly in the company's own terminology, and always asks for your details in the same order. This is often achieved by fine-tuning a general model on thousands of earlier, successful customer service conversations.

Who is affected

developerIt must decide whether a good prompt and search are sufficient for a task, or if the cost of collecting examples and training is justified. → Code

entrepreneur, executiveA model fine-tuned on proprietary company data can be a competitive advantage, but it costs money, data work, and continuous maintenance.

employeeSample answers are often collected from the work of experienced colleagues, so your expertise can end up in the model too. → Work & HR

lawyer, citizenIf the training examples contain personal or copyright-protected data, these can leave traces in the model, raising data protection and legal issues. → Law

A common misconception

MythFine-tuning can reliably inject new knowledge and facts into a model, like into a database.

In realitySFT primarily teaches behavior, style, and problem-solving patterns. For precise, frequently changing facts, it is usually better if the model retrieves data from an external source before answering.

Limits and risks

The model will be only as good as the examples: it learns faulty, one-sided, or biased sample answers just the same.
With too few or too similar examples, the model memorizes and performs worse in new situations. Meanwhile, it can also lose some of its previous general capabilities.
Fine-tuning can weaken a model's built-in safety limits, so it must be rechecked afterward for how it behaves on sensitive requests.
When a new base model appears, training usually needs to be redone on it, but the dataset is often reusable.

Similar, but not the same

pre-training – Pre-training teaches the language from massive, unlabeled text, while SFT builds on this by specializing the model with a small, carefully curated dataset.
reinforcement learning – There, the model does not receive a ready sample answer but tries, and a reward or score indicates how good the result was.
retrieval-augmented generation (RAG) – RAG does not change the model; it simply places retrieved documents next to the question, whereas SFT rewrites the model itself.

Related concepts

Broader

fine-tuning, supervised learning, machine learning, transfer learning

Narrower

instruction-based fine-tuning

On the edge

pre-training, reinforcement learning with human feedback (RLHF), parameter-efficient fine-tuning, LoRA, knowledge distillation, training data, prompt

A brief history

Supervised learning is the classic foundational method of machine learning. For language models, the recipe of "pre-training first, then fine-tuning" spread in 2018, following the ULMFiT method, as well as the first GPT and BERT models.

Where things stand

As of late 2026, SFT is a standard step in building language models. Several major providers offer it to companies as a ready-made service, but not for every model. However, for many tasks a good prompt and retrieval are a cheaper solution, so it is mainly used where style, format, or a narrow specialist domain matters.

Status: 2026 November

Also known as

SFT · supervised fine-tuning · supervised fine-tuning · supervised fine-tuning (SFT) · supervised finetuning · supervised fine-tuning

Sources

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.