How many examples do we actually need?

Far fewer than building from nothing, often hundreds rather than tens of thousands, and the figure depends on how close your task is to what the starting model already learnt. The reliable way to find out is trying it with what you have, since the answer for your situation is cheaper to measure than to estimate.

Does it damage what the model was good at?

It can, if you adjust too much of it on a narrow set, which is the same erosion that makes a heavily fine-tuned model worse at things nobody meant to change. Adjusting the smallest part that gets the result, and testing on tasks you were not targeting, is what keeps that in check.

Is this what a vendor means by training on our data?

Sometimes, and the phrase covers several different things, so it is worth asking which. Adapting an existing model on your examples is this. Putting your documents in front of a model at the moment of asking is not training at all. The two have completely different implications for cost, privacy and how updates work.

Glossary

Transfer learning

Transfer learning means reusing what a model learnt on one large task as the starting point for a different and usually smaller one, rather than building something from nothing each time.

In plain terms

Hiring somebody experienced rather than training a school leaver. What a model learnt from an enormous general task, language, or images, or sound, turns out to be most of what it needs for a narrower job, so somebody starting a new project begins from that rather than from a blank system. The remaining work is small enough for an ordinary organisation to do.

Why it matters

It is the reason a small company can build anything in this field at all. Before it became standard, a system that recognised your particular defect needed tens of thousands of labelled examples and a research team; now it needs hundreds and an afternoon of training, because the general understanding of shapes and edges was already there. Almost every practical AI project you will encounter rests on this without mentioning it.

How it works

You start from a model already trained on a large general task and continue training it on your smaller, specific one. The general knowledge is retained and adapted rather than replaced, which is where the saving comes from and why the amount of your own material needed drops so sharply.

How much of the original you leave untouched is the main decision. Adjusting only the last part is quick, needs least data and preserves general ability; adjusting more can fit an unusual task better and risks eroding what made the starting point valuable. Starting with the smallest adjustment and increasing only if results demand it is the ordinary practice.

It works best where your task genuinely resembles the original one. Adapting a general image model to spot defects on a production line is close enough that it works; adapting it to interpret ultrasound scans is further away and needs considerably more. How far apart the two tasks are predicts effort better than the size of either.

The same principle sits underneath fine-tuning a language model, which is the version most business readers meet. That page describes the practicalities and this one describes why it is possible at all, which is that the expensive general learning was done once and can be borrowed indefinitely.

The same defect detector, built two ways

The difference between these columns is not effort but eligibility. Building from nothing was a research undertaking, and the organisations that could contemplate it were the ones with research departments. Starting from a model that already understands edges, shapes and textures leaves you with a much smaller question, which is what your particular defect looks like, and that question can be answered with a few hundred photographs and a few weeks. Almost every practical project in this field now works that way, usually without anybody naming it. That is worth knowing when a quote arrives: the number is small because the expensive part was done once, elsewhere, and is being borrowed rather than repeated.

Seen in the wild

Adapt a published general model on a few hundred of your own examples and notice how much less material is required than building from nothing would need.
Ollama
Compare a general assistant against a narrow adapted model on the same task through one interface, which shows what the adaptation bought.
OpenRouter
Try the cheapest version first by putting examples in the request rather than training anything, which frequently closes enough of the gap.
Claude

Common misconceptions

People assume

It only applies to fine-tuning language models.

In fact

It is the general principle behind most practical AI work, including nearly every image project. The language version is simply the one business readers meet by name. Understanding it as a principle explains why so many projects that sound enormous turn out to need a few hundred examples.

People assume

Any starting point will do.

In fact

How closely the original task resembles yours decides how much you save. A general image model adapts readily to another everyday visual task and far less readily to medical scans, which look nothing like what it learnt from. Choosing a starting point near your problem is worth more than choosing a larger one.

Telling them apart