What is a dataset? A collection of examples gathered to teach or test an AI

A collection of examples gathered to teach or test an AI.

A dataset is a set of examples put together for an AI: photos with labels, sentences with translations, or huge piles of text collected from the web. It's usually split into a part for training and a part kept back for testing, so the model is checked on questions it hasn't seen.

An AI can only learn what's in its data. If the data leans one way, the answers lean too; if it contains mistakes, the model learns the mistakes. So what was collected, where from, and with whose permission matter as much as how much. ImageNet, released in 2009, is often cited as proof that a dataset can lift a whole field.

Think of it like this

A dataset is an AI's workbook. If the workbook only has problems from one chapter, you only get good at that chapter. And if the test questions are already in the workbook, the test measures memory, not skill.

For example

MNIST, a set of 70,000 photos of handwritten digits, is one of the most famous practice datasets.

Common mix-ups

Mix-upAs long as there's lots of data, you get a good AI.
ActuallyThe Pile, a text collection released in 2020, gathered 825 GiB from 22 different sources for variety, yet a book collection inside it, Books3, later left the models trained on it carrying copyright problems. Size and variety alone aren't enough.
Mix-upPutting the test questions into the training data shows real progress.
ActuallyEven MNIST, a standard test bed since 1998, was released with its training and test pictures split from the start. Researchers make it a basic rule to keep the test set locked away, unopened, until the very end.

Try it in class

For teachers at the start of a lesson — or for students working on their own.

Question: What happens if you teach "recognise a dog" with photos of only one breed?
Activity: In groups, plan a "fruit photo dataset". Write down which fruit, how many photos of each, where they'll come from, and how you'll split them into training and test sets. Then open the MNIST (E0017), ImageNet (E0022) and The Pile (E0113) records and compare when real datasets came out and how big they were.