scikit-learn

House prices from California in 1990, broken down by neighborhood with information about income, population, and how old the houses are. It's a go-to dataset for teaching machine learning because it has enough real-world messiness (missing values, outliers, geographic variation) to be useful without being overwhelming. If you've seen a tutorial showing how to predict prices or identify patterns in data, there's a good chance it used this dataset.

Imagine a collection of 1,797 tiny pictures of handwritten numbers (0 through 9), each one just 8 pixels by 8 pixels in size. Every picture is labeled with the correct digit. This dataset is what you'd use if you wanted to teach a computer to recognize handwriting. It's smaller and faster to work with than the famous MNIST dataset (which has way more images), so it's great for learning machine learning without needing a supercomputer.

A dataset from the 1970s showing home prices in different Boston neighborhoods alongside details about those neighborhoods. For each area, you can see things like crime rates, how many rooms homes have on average, what percentage of residents are lower-income, and how close the area is to jobs. It's been used a lot in data science classes to teach people how to build models that predict house prices based on neighborhood features. Researchers discovered later that one of the variables had a problematic history, which is why newer textbooks use different datasets instead.

Imagine you're a doctor trying to figure out which health measurements today predict how fast a patient's diabetes will get worse over the next year. This dataset has exactly that: 442 real patients, their 10 basic health numbers at the start (things like age, weight, cholesterol), and how much their diabetes progressed after a year. It's perfect for testing prediction models or learning which measurements actually matter most.

Imagine you have three different wines from Italy and you want to tell them apart. This dataset shows you 13 different measurements you could take from each bottle: how much alcohol it has, how acidic it is, how colored it looks, what compounds are in it. Scientists did this for 178 bottles, so now you have 178 examples where you know both the measurements and which type of wine it actually was. It's like having an answer key. People use this to test out new ways to identify wines from their chemical properties, or to practice building algorithms that can sort things into categories.