What is a dataset?
In one line
Ages 8 to 10: A dataset is a big, tidy collection of examples, like a sticker album where every sticker has the same facts written underneath.
Ages 11 to 13: In a dataset, each row is one example and each column is one feature or the label, and the rows are usually split into training and test sets.
Ages 14 to 18: A dataset carries choices about who and what was collected, so good practice documents its source, gaps and intended use alongside the data.
A dataset is an organized collection of examples gathered for a purpose, such as training or testing an AI model. Each example is one row, like one photo or one message. Each column records one fact about it, such as its size, its color or the right answer.
Last reviewed: 20 October 2026 6 min read
The idea in one picture
Think of a class survey sheet. Every student fills in one line, and every line has the same boxes. A dataset is that sheet, made tidy enough for a computer. Here is a small one about fruit:
| Weight (g) | Color | Label |
|---|---|---|
| 150 | red | apple |
| 170 | green | apple |
| 120 | yellow | banana |
| 130 | yellow | banana |
| 160 | red | apple |
| 115 | yellow | banana |
It has 6 rows, so 6 examples. Each row has two features, weight and color, and one label, the right answer. Before any learning, the last two rows are put aside as a test set. A model learns from the first four and is then checked on the two it never saw.
How it works, step by step
- Choose the question. For example: is this fruit an apple or a banana?
- Collect examples. Each one becomes a row.
- Choose the columns. Decide which facts, or features, to record for every example.
- Add labels. For supervised learning, each row gets its right answer.
- Clean it. Fix typing mistakes, remove duplicates and decide what to do about missing values.
- Split it. Keep a test set apart before training starts.
A little deeper (ages 11 to 13)
Many projects split the data three ways. The training set teaches the model. A validation set is used to compare settings and choose the best. The test set is used once, at the very end, to estimate how well the model works on new cases.
Size is not everything. A dataset of a million photos taken only in daylight teaches little about night. What matters is whether the examples cover the cases the model will meet. Labels can be wrong, too: researchers have found mislabeled examples even in the test sets of well-known datasets, which can change which model looks best.
Going further (ages 14 to 18)
A dataset is a set of decisions: what to collect, from whom, how to label it and what to leave out. One proposal is that every dataset should come with a datasheet, a short document describing why it was made, what it contains, how it was collected and what uses it is recommended for.
When a dataset does not represent the people or cases a system will serve, the system can be less accurate for the groups that are missing. And if any information from the test set slips into training or tuning, the test score stops telling you how the model does on new data. This is called leakage. A dataset about people also raises privacy questions: who agreed to be included, and what is stored about them?
Common mix-ups
- Mix-up: "A dataset and training data are the same."Fact: Training data is the part of a dataset used to teach the model. The rest is kept for validation and testing.
- Mix-up: "Bigger is always better."Fact: A huge dataset with the wrong examples can be worse than a smaller one that covers the right cases.
- Mix-up: "Data is just facts."Fact: Labels and columns are choices made by people, and labels can be wrong.
- Mix-up: "Once collected, a dataset is finished."Fact: Good datasets are documented, checked and updated, and their limits are written down.
Try it (no device, 5 minutes)
On paper, make a dataset of 8 objects in the room. Use three columns: color, "heavier than a book?" (yes or no) and a label, "toy" or "not toy". Cover the last two rows. Ask a partner to study the first six and guess the labels of the covered two from their color and weight. Were they right? Which column helped most? What extra column would help?
Quick check
In the fruit table, how many examples and how many features are there?
Answer: 6 examples (rows) and 2 features (weight and color). The label column is the answer, not a feature.
6 examples (rows) and 2 features (weight and color). The label column is the answer, not a feature.Why are some rows kept aside before training?
Answer: So the model can be tested on examples it has never seen.
So the model can be tested on examples it has never seen.What should a datasheet for a dataset tell you?
Answer: Why the dataset was made, what is in it, how it was collected, and what uses it is recommended for.
Why the dataset was made, what is in it, how it was collected, and what uses it is recommended for.Questions
A dataset is a tidy collection of examples, where every example has the same kinds of facts recorded about it. In a table, each row is one example and each column is one fact, such as a size, a color or the right answer.
It is the collection of examples a model learns from and is tested on. The dataset is usually split into a training set for learning and a test set for checking the model on examples it has not seen.
A table of fruits with their weight, color and name is a small dataset. Larger ones include collections of labeled photos, recorded speech with its written text, or messages marked as spam or not spam.
A good dataset covers the cases the model will meet, has accurate labels and records where the data came from. It also keeps test examples separate from training examples and comes with notes on what it is suitable for.
A dataset is the whole collection. Training data is the part of it used to teach the model; other parts are kept for validation and testing.
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.