Fine-Tuning
Fine-tuning is the process of taking an AI model that has already been trained on broad, general data and training it a bit further on a smaller, focused set of examples, so it specializes in a particular task, domain, or style.
What is Fine-Tuning?
Training a capable AI model from scratch is enormously expensive — it takes vast amounts of data, computing power, and time that almost no one can afford. Fine-tuning is the shortcut that sidesteps most of that cost. Instead of starting from nothing, you begin with a model that has already done the hard, general learning — a model that already understands language, or images, broadly — and you give it a comparatively small amount of additional, targeted training on examples from your specific area. The model doesn't relearn everything; it adjusts what it already knows to fit the narrower job you actually care about. The result is a specialist built on top of a generalist, for a tiny fraction of the effort it would take to build that specialist from the ground up.
What fine-tuning changes is the model itself: the internal values that shape its behavior — its parameters, or weights — shift to reflect the new examples. That makes it the right tool when you want to change how a model behaves: to adopt a consistent house style, to handle the specialized vocabulary of a field, to reliably produce output in a particular format, or to get better at a task that the general model only does so-so. It is worth being clear about what fine-tuning is not, because the two are often confused. If your goal is simply to give a model access to current or private facts, retrieval-augmented generation — feeding the model the right reference documents at the moment you ask — is usually the better and cheaper choice. The rough rule: fine-tuning teaches a model new behavior and skills; retrieval supplies it with new facts. Plenty of real systems use both together.
Fine-tuning comes in several flavors. Full fine-tuning adjusts the entire model and is the most thorough but also the most demanding. More popular today are lighter, cheaper methods — often grouped under the label parameter-efficient fine-tuning (PEFT), with low-rank adaptation (LoRA) being a well-known example — that tweak only a small slice of the model while leaving the bulk of it frozen. These get most of the benefit for a fraction of the cost, which is a big part of why fine-tuning has become something small teams and even individuals can do, rather than something reserved for the labs that build the original models.
Real-world example
A radiology group keeps getting frustrated with off-the-shelf voice transcription: a general speech-to-text model mishears dense medical terms, turning "pneumothorax" into nonsense and forcing doctors to correct every report by hand. So they fine-tune the model on thousands of hours of their own radiologists' dictations, paired with the correct transcripts. The model doesn't have to relearn English — it already knows that. It just adjusts to the specialized vocabulary, the pacing, and the phrasing of radiology dictation. Afterward it transcribes their reports far more accurately, because it has been nudged from a capable generalist into a specialist tuned to exactly how their doctors actually speak.
Related terms
Frequently asked questions
What is the difference between fine-tuning and RAG?
They solve different problems. Fine-tuning trains the model further so it changes how it behaves — its style, its format, its skill at a particular task. Retrieval-augmented generation (RAG) leaves the model unchanged and instead hands it the right reference material at the moment you ask a question. So if you want the model to write like your team, fine-tuning helps; if you want it to know your latest documents, RAG helps. They are complementary, and many production systems combine the two.
How much data do you need to fine-tune a model?
Far less than training one from scratch, but the honest answer is "it depends" — on the task, the base model, and how different your target is from what the model already does. Teaching a familiar model a consistent style or format can take a surprisingly modest number of good examples, while adapting it to a genuinely specialized domain needs more. The bigger lever is usually quality over quantity: a smaller set of clean, well-chosen, representative examples typically beats a larger pile of noisy ones.
Do I need to fine-tune a model to use AI in my business?
Usually not, at least at first. For most needs, a good general model combined with clear prompting and, where relevant, retrieval over your own documents will get you a long way without the cost and upkeep of fine-tuning. Fine-tuning earns its keep when you need consistent behavior the general model can't reliably deliver — a specific voice, a niche vocabulary, a fixed output format at scale. It's best thought of as a step you reach for when simpler approaches fall short, not the default starting point.