Training Dataset

Modern AI programs are not written rule by rule by humans. Instead, they are shown a very large number of examples and left to derive patterns from them. This collection of examples is called a training dataset. For a program meant to recognize cats in photos, it consists of millions of images along with an indication of what is shown in each one. For a chat program, it consists of enormous amounts of text: websites, books, forums, program code. The dataset is therefore not part of the finished product, but the material from which it emerges.

The dataset sets the upper limit

A program can only learn patterns that appear in its examples. Anyone who trains an image-recognition system exclusively on photos from Europe ends up with a system that performs worse outside Europe. This is not a programming error but a property of the data. That is why it is said: the training dataset sets the upper limit for what is ultimately possible.

Particularly tricky are distortions that no one built in on purpose. If application data spanning thirty years predominantly features male executives, a model learns this correlation along with everything else. It treats it as a pattern of the world rather than a pattern of the past. Experts call this kind of thing bias, meaning a systematic skew. It can later only be corrected out with great effort.

On top of that, there is a legal dimension. Many large text collections originate from the open internet and contain copyrighted works. Publishers, photo agencies, and authors have therefore been suing AI providers for several years now. For investors, this is a genuine business risk, because it is difficult to subsequently purge a model of the influence of individual pieces of data.

From raw material to a usable dataset

It begins with collection. Programs automatically crawl websites and save their content, or a company purchases ready-made collections. What follows is the most labor-intensive part: cleaning up. Duplicate texts are removed, advertising junk and broken pages are sorted out, and often certain content is deliberately filtered out as well.

For many tasks, each example additionally needs an answer that the model is supposed to produce. This extra piece of information is called a label. A photo gets the label “dog,” a customer comment the label “negative.” Labels are often assigned by humans by hand, which is expensive and slow. With language models, this problem is sidestepped: the model learns to predict the next word of a text, and the text itself supplies the correct answer.

An important trick is the split. A portion of the data is held back and never shown during training. This test data is later used to check whether the model has truly understood something. Otherwise, it might simply have memorized the training examples. This memorization is called overfitting and is one of the most common mistakes made during training.

When datasets make the news

In product announcements, training data usually shows up as a number. Providers cite, for instance, the amount of text, measured in tokens — that is, in units of words and word fragments. Figures in the range of several trillion tokens are by now normal for large language models. If, on the other hand, a provider leaves open where its data comes from, that omission is itself the interesting piece of information.

This topic also comes up in everyday life, usually in the fine print. When a service asks whether it may use your chats to improve the model, it is talking about future training data. That is why many companies explicitly promise business customers not to use their inputs for this purpose. For companies handling customer data, this is often the decisive condition.

A common misconception is that a running model continually keeps learning from conversations. This is normally not the case. Training is a completed process, and afterward the model is fixed. New knowledge is either added through a later round of training or supplied along with the request itself, for instance from a search.