fine-tuning

basicwidespread

Fine-tuning is the further training of an already trained AI model on specific, targeted examples to better handle a particular task, style, or field.

In plain terms

Imagine hiring a freshly graduated, well-educated doctor who then takes a specialist exam in dermatology at your hospital. They do not retake university: building on existing knowledge, they learn a narrower field based on many practical cases. Fine-tuning does the same with an AI model.

Large language models are trained on vast amounts of general text, which costs a lot of money and time. In fine-tuning, this pre-trained model is further trained on a much smaller, carefully selected dataset. This way, it can learn, for example, a company's tone, a profession's vocabulary, or a specific response format.

What makes it different from a simple instruction (prompt) is that the model itself changes permanently here. You do not need to explain what you expect from it with every question, because the learned behavior is embedded. In exchange, creating it requires more work, time, and money, and if the examples are poor, the model learns the mistakes too.

How it works

As a starting point, a pre-trained base model is selected that generally handles language, images, or code well.
A set of examples is compiled for the target task, typically a few hundred or a few thousand input–expected output pairs, such as a question and a model answer to it.
The model attempts to solve the examples one by one, and based on the discrepancies, its internal numerical values (weights) are adjusted slightly to get closer to the expected answer next time.
Often the entire model is not rewritten: the original weights are left unchanged, and only a small number of additional parameters are trained on top of them (the LoRA method is an example), at a fraction of the cost.
Finally, they test on examples not seen during training and verify whether general knowledge has degraded in the process.

An everyday example

You call your insurer's customer service, and a chatbot answers, speaking in exactly the company's terminology, in the usual polite style, and using its own product names. A general language model may be behind it, fine-tuned on the company's earlier customer conversations, anonymized beforehand, although the same can often be achieved with a good prompt and by connecting the company's documents.

Who is affected

entrepreneur, executiveYou must decide whether a model further trained on your own data is worth it, or if a good prompt plus linking company documents suffices.

developerYou assemble the training examples, you run the training, and you measure whether the model has truly improved. → Code

employeeYour workplace AI tool may know your profession's terminology and your company's habits because it was fine-tuned, potentially using examples from your own work. → Work & HR

patient, doctorMany systems that summarize medical reports or analyze images are variants of a general model further trained on specialized data, so their quality largely depends on the training data. → Health

A common misconception

MythFine-tuning allows new knowledge and facts to be loaded into the model, like into a database.

In realityFine-tuning primarily teaches behavior, style, and problem-solving patterns, but it records facts unreliably. If up-to-date data is needed, a better solution is for the model to search documents when answering.

Limits and risks

The model is only as good as its examples: flawed, one-sided, or biased training data yields a flawed and biased model.
A model trained for a narrow task may forget things it previously knew, so its general capabilities must be retested.
Personal or business secrets present in the training data can become embedded in the model, making them difficult or impossible to remove later.
A fine-tuned model does not update itself: if regulations, the product, or the base model change, it must be retrained.

Similar, but not the same

prompting (instruction writing) – With prompting, you only phrase the request more skillfully while the model remains unchanged; with fine-tuning, the model itself is altered.
retrieval-augmented generation (RAG) – RAG retrieves fresh information from external documents when answering, while fine-tuning permanently embeds learned knowledge into the model in advance.
pre-training – Pre-training is the model's first, general training on a massive amount of data, while fine-tuning is the subsequent, smaller, and targeted second step.

Related concepts

Broader

machine learning, deep learning, transfer learning, model training

Narrower

LoRA, parameter-efficient fine-tuning, instruction-based fine-tuning, reinforcement fine-tuning

On the edge

foundation model, large language model, pre-training, retrieval-augmented generation, training data

A brief history

The method grew out of the concept of transfer learning: knowledge acquired elsewhere is valuable for new tasks. In image recognition, it became standard practice in the early 2010s to fine-tune a network pre-trained on a large image collection for new tasks. For language models, this became general practice in 2018: ULMFiT, OpenAI's first GPT model, and especially Google's BERT model demonstrated that a pre-trained model could be adapted to many tasks with few examples. The LoRA method, published in 2021, significantly reduced the cost of fine-tuning large models by requiring training of only a few additional parameters.

Where things stand

By the end of 2026, fine-tuning is a standard procedure: major providers offer it as a service, and open models can also be further trained on your own machines. For many tasks, a good prompt and document retrieval are enough, so it is mainly used where a consistent style, specialized expertise, or a smaller, cheaper model is needed.

Status: 2026 November

In the news

Appears in 1 Agent Age stories; the curve shows weekly occurrence in the news magazine.

European Commission draft would regulate workplace AI use to protect workersOct 11, 2026 · Jog

Also known as

fine-tuning · fine-tuning · fine tuning · finetuning · finomhangolt modell · model fine-tuning · further training

Sources

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