What is training data?
In one line
Ages 8 to 10: Training data is the practice examples a machine learns from, so if those examples are odd or leave something out, the machine's answers will be too.
Ages 11 to 13: Training data is the part of a dataset used to set the model's numbers, while a separate test set checks how well it works on new cases.
Ages 14 to 18: The coverage, labels and origin of training data limit what a model can do, and any overlap with test data can make results look better than they are.
Training data is the set of examples an AI model learns from. The model adjusts itself until its answers fit these examples, so whatever is in them (good patterns, mistakes and gaps) shapes what it does next. It is kept separate from test data, which checks the model on examples it has not seen.
Last reviewed: 20 October 2026 6 min read
The idea in one picture
Imagine teaching a recycling sorter with 40 labeled photos: 30 cans and 10 bottles. Every bottle in the photos happens to be clear glass. The sorter learns well and gets every training photo right.
Then a green bottle arrives. The sorter has never seen a green bottle, and it may call it a can, or anything else, because nothing in its training data showed it what a green bottle looks like. The sorter did exactly what its examples taught it. The gap was in the training data.
The fix is not a cleverer sorter. It is better training data: add labeled photos of green, brown and plastic bottles, then test again on photos the sorter has never seen.
How it works, step by step
- Collect examples of the cases the model should handle.
- Label them with the right answers, if the task is supervised learning.
- Split off a test set before any training starts.
- Train. The model adjusts its parameters, its internal numbers, to fit the training examples.
- Keep the parameters. After training, a neural network or similar model runs on its learned numbers; it does not need the training examples to make a prediction.
- Test on the held-back examples and count the mistakes.
A little deeper (ages 11 to 13)
Three things decide whether training data is good:
- Coverage. Does it include the kinds of case the model will meet: different lighting, accents, handwriting, ages?
- Labels. Are the answers right? Researchers have found mislabeled examples even in the test sets of widely used datasets.
- Size. More examples usually help, but only if they add variety. A thousand copies of the same clear bottle teach little.
Projects often use three parts of a dataset: a training set to learn from, a validation set to choose settings, and a test set used once at the end.
Going further (ages 14 to 18)
Training data limits a model in ways a score can hide. When some groups are under-represented, accuracy for those groups can be lower. When test examples, or information about them, slip into training or tuning, the test score stops reporting how the model does on new data. This is leakage, and it is easy to cause by accident, for example by preparing all the data together before splitting it.
Where training data comes from matters too. Documentation such as a datasheet records why it was collected, what it contains and how it was gathered. A study also showed that a large language model could reproduce exact text from its training data, including personal information, which is one reason training data about people raises privacy questions.
Common mix-ups
- Mix-up: "Training data and test data are the same."Fact: They must be kept apart, or the test proves nothing.
- Mix-up: "The model looks things up in its training data."Fact: A neural network uses its learned parameters. The examples shaped those numbers but are not consulted when it answers.
- Mix-up: "More data always fixes it."Fact: More of the same examples does not fill a gap. Variety and correct labels matter.
- Mix-up: "Training data is neutral."Fact: It reflects choices about what to collect and how to label it.
Try it (no device, 5 minutes)
Draw six trees, all the same kind: a round green top on a brown trunk. Show them to a partner as "examples of trees". Now draw a pine tree, a palm tree and a bare winter tree, and ask: would a machine that learned only from your six drawings call these trees? What three drawings would you add to the training data?
Quick check
Why did the recycling sorter struggle with the green bottle?
Answer: Its training data had no green bottles, so it never learned what one looks like.
Its training data had no green bottles, so it never learned what one looks like.Why must the test set be kept apart from the training data?
Answer: If the model has already seen the test examples, a good score shows only that it remembers them, not that it works on new cases.
If the model has already seen the test examples, a good score shows only that it remembers them, not that it works on new cases.What is leakage?
Answer: When information from the test set gets into training or tuning, so the test score can look better than the model's real performance on new data.
When information from the test set gets into training or tuning, so the test score can look better than the model's real performance on new data.Questions
Training data is the collection of examples an AI model learns from. The model adjusts itself to fit these examples, so what is in them, and what is missing, shapes how it behaves.
Training data is used to teach the model. Test data is kept apart and used afterwards to check how well the model works on examples it has never seen.
A model can only learn patterns that appear in its training data. Gaps, wrong labels and unfair patterns in the data can turn into mistakes in the model.
Photos labeled with what they show, recordings of speech with their written words, and messages marked as spam or not spam are all training data. So is a table of past house sales with their prices.
A neural network stores learned numbers called parameters, not a filing cabinet of its examples, and it does not look them up to answer. Even so, a study showed that a large language model could reproduce exact text from its training data, including personal information, so data about people needs care.
Mechna Academy is a city of machines for curious children aged 8 to 18, opening first to ages 8 to 10. In each mission, a child guesses how a machine decides, then opens its Peek Hatch to check, changes one thing and tests it on a case it has never seen. Children learn how AI works this way, and they never chat with an AI.
For grown-ups: join the waitlistPlaces are limited. The waitlist is for parents and guardians 18 and older.