Dataset is a structured collection of data used for training, validating, or testing a machine learning model. It is the raw material from which models learn and against which they are measured.
What it means in plain English
A dataset is simply an organised collection of examples — rows of records, a folder of images, a body of text. In machine learning, datasets are usually split into parts: a training set the model learns from, and a separate test set used to check how well it performs on data it has not seen. Keeping these separate is essential for an honest measure of a model’s ability.
The scale of datasets varies enormously, from a few hundred rows for a simple task to trillions of words for a large language model.
A simple example
A dataset of house sales — each with features like size, location, and price — could be used to train a model to predict prices. Part of it trains the model; a held-back part tests whether its predictions hold up on unseen sales.
Why it matters
Datasets are the foundation of all machine learning. The availability of large, high-quality datasets is one of the main reasons AI has advanced so rapidly, and how a dataset is built and split directly affects how trustworthy a model’s results are.
Related terms
- Training Data — the portion of a dataset used to teach a model.
- Model — what a dataset is used to build.
- Feature — the individual measurable properties within a dataset.
Frequently asked questions
What makes a good dataset?
Quality, quantity, and representativeness: accurate data, enough of it, and coverage of the real situations the model will face. A biased or too-small dataset limits how well a model can perform.
What are training, validation, and test sets?
A dataset is usually split three ways: the training set teaches the model, the validation set tunes it, and the test set gives an unbiased final measure of how well it generalises to unseen data.
