Adult Census Income Benchmark

## Overview The Adult dataset (a.k.a. Census Income) is one of the most cited tabular benchmarks in machine learning. It contains ~48,842 rows and 15 columns (14 features + binary target) extracted by Barry Becker from the 1994 US Census Bureau database. The task is to predict whether an individual's annual income exceeds $50,000 based on demographic and employment attributes. Format: ARFF (OpenML v1); easily convertible to CSV/Parquet. ## Schema - age — numeric — age in years - workclass — categorical — employment type (Private, Self-emp, Gov, etc.) - fnlwgt — numeric — census final weight - education — categorical — highest education level - education-num — numeric — education level encoded as integer - marital-status — categorical — marital status - occupation — categorical — occupation category - relationship — categorical — household relationship - race — categorical — race (5 levels) - sex — categorical — Male/Female - capital-gain — numeric — capital gains - capital-loss — numeric — capital losses - hours-per-week — numeric — typical work hours per week - native-country — categorical — country of origin - class — binary target — >50K or <=50K ## Sources - OpenML dataset 179 (Adult): https://www.openml.org/d/179 — Public Domain - Original: UCI Machine Learning Repository (Adult), donated 1996 by Ronny Kohavi & Barry Becker: https://archive.ics.uci.edu/dataset/2/adult ## Methodology Extracted from the 1994 US Census Bureau database by Barry Becker using the following filtering conditions: ((AAGE>16) && (AGI>100) && (AFNLWGT>1) && (HRSWK>0)). OpenML hosts the standardized ARFF version with consistent attribute typing. Missing values are encoded as '?' in workclass, occupation, and native-country. ## Known gaps & limitations Data reflects 1994 US demographics and income distributions — not representative of current populations or economies. Known class imbalance (~76% <=50K). Documented demographic biases (race, sex, native-country distributions skew toward US-born white males) make this dataset a frequent case study in algorithmic fairness literature; using it for production income prediction is inappropriate. Contains missing values flagged as '?'. The fnlwgt column is a survey weight and is generally not meaningful as a predictive feature. ## Intended use & out-of-scope - IS for: tabular ML benchmarking, AutoML evaluation, fairness/bias research, teaching examples, baseline comparisons for classification algorithms. - NOT for: real-world income prediction, contemporary demographic analysis, or any deployed decision-making about individuals. Heavily present in ML training corpora — leakage risk if used to evaluate foundation models on tabular tasks. Original supplier listing: Adult (Census Income) — UCI/OpenML Benchmark Classic UCI 'Adult' census income dataset (~48K rows, 14 features) for predicting whether income exceeds $50K/yr. Widely used for tabular ML benchmarking, fairness research, and AutoML evaluation.