UC Irvine Machine Learning Repository

A bank in Portugal called 45,000+ customers to sell them term deposits (savings products). This data captures who they called, what they said on the phone, and whether each person agreed to sign up. You get details like each caller's age, job, how long the conversation lasted, and whether they already had loans or credit issues. It's a real-world example of how banks do sales outreach and what works.

Unlike modern income surveys that track trends year-over-year, this dataset is frozen in time: it's a snapshot of about 49,000 working-age people from 1994, with their age, job, education level, and whether they made more or less than $50,000 that year. It's not meant to show what salaries look like today. Instead, it's become famous among data scientists and students as a training ground for learning how to build predictive models. You feed the computer someone's demographics and work hours, and it learns to guess whether that person's in the higher or lower income bracket. It's been used in thousands of machine learning courses and research papers.

Want to know what makes a good white wine scientifically? This dataset has lab test results from nearly 5,000 Portuguese white wines, paired with expert quality scores. You get measurements like alcohol content, acidity, sugar, sulfites, and density for each bottle, plus ratings from 0 to 10. It's the kind of data that lets you explore whether certain chemicals predict whether a wine tastes good or bad. Great if you're curious about wine chemistry or building models to guess quality from lab numbers alone.

Red wine samples tested for their chemical makeup and then scored by experts on quality. You get measurements like how much alcohol, acid, and sulfur dioxide are in each wine, plus the final quality rating it received. It's useful if you're curious about what makes wines taste better or want to build a system that predicts wine quality from lab results. This is popular with data scientists learning machine learning because it's small enough to work with quickly but has enough real variation to be interesting.

Doctors took tissue samples from breast masses and measured their physical characteristics using microscopes and image analysis. This dataset contains 30 different measurements for each of 569 samples, things like how round the cells are, how tightly packed together they are, and how bumpy their edges look. Each sample is labeled as either cancerous (malignant) or not cancerous (benign). It's been used thousands of times by researchers testing new ways to automatically diagnose cancer from cell images, and it's popular because the measurements are straightforward and the results are clear.

303 people from a Cleveland clinic had their heart health checked, and doctors recorded 13 measurements like age, cholesterol, blood pressure, and how their heart responded to exercise. The goal is simple: can you predict whether someone has heart disease just from these numbers? It's one of the most popular practice datasets for this type of prediction problem.