
Fine-Tuning
Fine-tuning means taking an already fully trained AI model and training it further with additional, specially selected examples. This adapts a general model to a specific task, language, or company style without building it from scratch.
Large AI programs first learn from enormous amounts of text from the internet. Afterward, they can do a little bit of everything, but nothing particularly precisely. Fine-tuning is the second step: the finished program is trained further, using a much smaller, carefully selected collection of examples. This turns an all-rounder into a specialist for a specific task. The comparison to human education fits well: first comes general schooling, then the apprenticeship at a specific company. Nobody starts back in first grade for the apprenticeship.
Why nobody starts from zero anymore
Building a large language model from scratch easily costs several million euros. It requires thousands of specialized chips computing for weeks, and vast amounts of text. Fine-tuning, by contrast, is cheap. Depending on the model size, a few hundred to a few thousand good examples and a few hours of computing time are enough.
This is exactly what has changed the AI market. A hospital, a law firm, or a software maker doesn’t need to develop its own model. It takes a freely available model and adapts it using its own data. That’s why companies like Meta or Mistral deliberately release their models for download: they provide the foundation, and others build their products on top of it.
Fine-tuning is also the reason chatbots answer politely instead of simply continuing text. The raw base model can only continue words. Only through further training with many question-answer pairs does it learn to behave like a conversational partner. Without this step, ChatGPT would never have become a usable product.
From example pairs to adjusted weights
A model consists of billions of numbers, so-called weights. They determine how the program gets from an input to an answer. During fine-tuning, the model is shown an example, its answer is compared to the desired answer, and the weights are shifted a tiny bit in the right direction. This is repeated thousands of times. The steps are deliberately small so that the old knowledge isn’t lost.
In practice, usually not all weights are changed. A popular method called LoRA trains only a few additional numbers while leaving the large remainder untouched. This saves enormous amounts of memory and cost. The result is a small additional file, often just a few megabytes in size, which is attached to the base model like an adapter.
The typical mistake is called overfitting: if you train too long on too few examples, the model just parrots those examples back. A related problem is forgetting previous abilities. A model that has been fine-tuned exclusively on legal texts may afterward answer everyday questions worse. That’s why general examples are often mixed in with the specialized data during training.
Fine-tuning in products and headlines
Anyone building AI into a product quickly encounters this decision. Providers like OpenAI or Google sell fine-tuning as a service: you upload your own example data and receive an adapted model in return. On the platform Hugging Face, tens of thousands of such adapted models are publicly available, for example for medicine, programming code, or individual languages.
It’s important to distinguish this from two neighboring concepts. A prompt is just an instruction in the chat window and doesn’t change the model at all. With the RAG method, the model is supplied with matching documents at runtime so it can look up facts. Fine-tuning, by contrast, changes the model itself. As a rule of thumb: for current factual knowledge, use RAG; for style, format, and behavior, use fine-tuning.
In business news, the term often appears in disputes over open models. If a company makes a model freely available for download, anyone can further train it, even toward goals the manufacturer doesn’t like. Safety rules can in fact be partially removed again through fine-tuning. This is precisely what regulators are currently discussing intensively.