What is training data, and why does it decide everything?

A cook learns from the food she has tasted. Give a cook twenty years at the coast and she will make excellent coastal food. Ask her for a highland stew and she will guess. She is not a bad cook. Her menu is narrow because her kitchen was.

Training data is the menu behind a model, and it is the pile of examples the model learned from. What the model does well, and what it gets wrong, both trace back to that pile.

This is why data work is most of the work in artificial intelligence. The clever part gets the headlines. However, the dull part decides the result.

What counts as training data

Training data is any set of examples a model is fitted to, and the form depends on the job.

• For a photo model, images with a label saying what is in them.
• For a speech model, audio clips with a written transcript.
• For a credit model, past loans with a note on whether they were repaid.
• For a language model, plain text on a very large scale.

Two words come up all the time, and a feature is a piece of information you feed in, such as age or rainfall. A label is the right answer for that example, and labels are the expensive part, because a person usually has to supply them. That loop is the heart of machine learning.

Where it comes from

Most data comes from four places. Records a company already holds, and public text and images from the web. Data bought from a vendor, and data made on purpose, by paying people to write, label or record.

Each source carries a catch, and company records match your business but are often thin. Web data is huge and full of junk and repeats. Bought data may be stale, and made data is clean, slow and costly.

There is a fifth source now. Models produce text and images, and that output is fed into the training of other models. It works up to a point. However, feed a model too much of its own kind of output and the quality drifts.

Why more data is not always better

Size helps, and size alone is not the goal. Three other things matter just as much.

The first is balance, and if nine in ten of your fraud examples come from one bank, the model learns that bank. The second is correctness, and a wrong label is worse than a missing one, because the model treats it as truth. The third is variety, and a thousand photos of the same maize field teach less than fifty fields in fifty places.

Repeats are a quiet problem too. The same page copied across many sites gets counted many times, and the model gives it too much weight. A neural network with billions of weights will memorize junk as readily as it memorizes truth. This is why serious teams spend weeks removing duplicates.

How bias gets in

Bias in a model is rarely put there on purpose, and it arrives with the examples.

Take hiring. If a firm promoted mostly one kind of person for twenty years, its records say that is what a good hire looks like. A model trained on those records copies the pattern, and it does not know the pattern was unfair. It only knows it was common.

Or take a voice assistant trained mostly on one accent. It will hear that accent well and struggle with others. Nobody wrote a rule against your accent, but the rule came out of the recordings.

Therefore, ask one question of any model you are offered. Who is in the data, and who is missing? The gaps predict the failures.

What happens at the edges

A model is reliable only inside the range of its examples, and outside that range it still answers, and it still sounds sure. Engineers call this being out of distribution.

A crop disease model trained in one country will misread a disease it has never seen. A large language model asked about a small town it barely read about will produce a fluent, wrong paragraph. Nothing warns you, because the model has no sense of where its own knowledge ends.

Language coverage, and why it is uneven

Text on the web is spread very unevenly across languages. A few languages have enormous amounts written down online, and most of the languages people speak have very little.

The effect is direct, and models write English well. They handle widely written languages such as Swahili or Hindi less well, and they handle small languages poorly. This is a data gap first and a technology gap second.

It is fixable, and the fix is dull work. Record speech, and write down stories, laws, textbooks and news. Get permission, and keep a record of it. Every hour of good, labeled local text is a small permanent asset.

Rights, consent and privacy

Data has owners, and text and images on the web belong to the people who made them. The rules on using that work for training differ by country, and courts are still arguing them out. Personal records carry stricter duties, and many countries now require a lawful reason to handle them at all.

Therefore, keep a written record for every data set. Where did it come from? What were we allowed to do with it? Who agreed, and when? A team that cannot answer those three questions has a problem waiting for it.

What to check before you trust a model

Ask the supplier four plain questions. What was it trained on? When did that data stop? Who is under represented in it? How was it tested, and on what?

Then run your own small test, and take fifty real cases from your own work, with answers you already know. Feed them in and count the mistakes. A GOOD MODEL ELSEWHERE CAN BE A BAD MODEL HERE. That test costs an afternoon, and it tells you more than any brochure.

Just Out Tech explains new research in plain language. This article was drafted with AI assistance and checked by a human against the original source.