The most famous dataset in machine learning. Measurements of sepal and petal length and width for three iris species (setosa, versicolor, virginica), 50 instances each. Introduced by R.A. Fisher in 1936.

The canonical toy dataset from the Weka book. Predicts whether conditions are suitable to play golf based on outlook, temperature, humidity, and wind. Used throughout the book to illustrate decision trees and Naïve Bayes.

All possible combinations of four nominal attributes for recommending soft, hard, or no contact lenses. Small but complete — no missing values.

Final settlements from Canadian labor negotiations (1987–1988). Mixed numeric and nominal attributes covering wages, hours, pension, and leave. Classifies contracts as acceptable or not.

Medical records from the National Institute of Diabetes for Pima Indian women aged 21+. Attributes include glucose concentration, blood pressure, BMI, insulin level, and diabetes pedigree function.

1984 U.S. House of Representatives voting records on 16 key issues. Each instance is a member of Congress classified as Democrat or Republican.

Chemical composition measurements for 214 glass samples from crime scene investigations. Six glass types including float/non-float window glass, containers, tableware, and headlamps.

Radar returns from a phased array of 16 HF antennas targeting free electrons in the ionosphere. Returns classified as "good" (evidence of structure) or "bad" (pass-throughs). 34 continuous attributes.

Each instance is a 3×3-pixel region drawn from seven outdoor images. 19 continuous attributes describe spectral and geometric properties. Seven classes: brickface, sky, foliage, cement, window, path, grass.

Extracted from the 1994 U.S. Census Bureau database. Predicts whether annual income exceeds $50,000 based on age, education, occupation, marital status, and hours worked. Widely used for bias and fairness research.

Credit risk classification for applicants at a German bank. Mixed attributes cover credit history, loan purpose, employment, and savings. Misclassifying a bad risk as good is five times more costly.

Patient records for diagnosing thyroid disorders. Seven continuous attributes (TSH, T3, TT4, T4U, FTI, age) and 23 nominal attributes. Four classes: negative, compensated, primary, and secondary hypothyroid.

Synthetic dataset forward-sampled from a 7-node Bayesian Network with a known causal structure: GeneticRisk + IrritantProducts → BrokenSkinBarrier; GeneticRisk + DustMiteExposure + HighSugarDiet → Th2Dysregulation; both → EczemaFlare. Designed for K2 structure learning — the true DAG is recoverable. ~25% positive (flare) class rate.