What is the Adult Income (Census) Dataset dataset?

Predict whether income exceeds $50K/yr based on 1994 US Census data. 32,561 records with 14 demographic, education, and employment features. The benchmark dataset for binary classification and algorithmic fairness research.

How can I download the Adult Income (Census) Dataset dataset?

You can download the Adult Income (Census) Dataset dataset in CSV or Parquet format directly from this page. Each table has its own download buttons.

What format is the Adult Income (Census) Dataset dataset available in?

This dataset is available in CSV and Parquet formats. CSV is great for spreadsheet applications, while Parquet is optimized for data analysis tools like Pandas and DuckDB.

What is the license for the Adult Income (Census) Dataset dataset?

This dataset is available under the CC BY 4.0 license. See the full license at https://creativecommons.org/licenses/by/4.0/

How many tables are in the Adult Income (Census) Dataset dataset?

The Adult Income (Census) Dataset dataset contains 1 table: adult_income.

Adult Income (Census) Dataset

Last updated 2 weeks ago•December 27, 2025

Source:UCI Machine Learning Repository

License:CC BY 4.0

Version:1.0

Time:1994

Location:United States

Created by Dataset Agent

Overview

The Adult Income Dataset (also known as Census Income Dataset) is the definitive benchmark for binary classification in machine learning. Extracted from the 1994 US Census Bureau database by Barry Becker and Ronny Kohavi, the prediction task is to determine whether an individual's annual income exceeds $50,000 based on demographic and employment attributes.

The dataset contains 32,561 instances with 14 attributes covering age, workclass, education, marital status, occupation, relationship, race, sex, capital gains/losses, hours worked, and native country.

View Source

SQL

SELECT COUNT(*) AS total_instances FROM adult_income.csv

Data

Total Instances
32,561
1 row

Originally donated to the UCI Machine Learning Repository in 1996, this dataset has become one of the most cited resources in machine learning literature—particularly for classification benchmarking, algorithmic fairness research, and studies examining income inequality and socioeconomic factors.

Dataset Characteristics

| Characteristic | Value | |----------------|-------| | Subject Area | Social Science | | Task Type | Binary Classification | | Feature Types | Categorical, Integer | | Instances | 32,561 (training) | | Features | 14 | | Missing Values | Yes (encoded as '?') |

The data was extracted using specific conditions from the Census database: ((AAGE>16) && (AGI>100) && (AFNLWGT>1) && (HRSWK>0)), ensuring only working-age individuals with positive income and work hours are included.

Class Distribution and Imbalance

The target variable shows significant class imbalance: 24,720 individuals (75.9%) earn ≤$50K while only 7,841 (24.1%) earn >$50K.

View Source

SQL

SELECT income, COUNT(*) AS count, ROUND( 100.0 * COUNT(*) / ( SELECT COUNT(*) FROM adult_income.csv ), 1 ) AS percentage FROM adult_income.csv GROUP BY income

Data

Income	Count	Percentage
<=50K	24,720	75.9
>50K	7,841	24.1
2 rows

View Source

SQL

SELECT income, COUNT(*) AS count FROM adult_income.csv GROUP BY income

Data

Income Level	Count
≤$50K (75.9%)	24,720
>$50K (24.1%)	7,841
2 rows

This ~3:1 class imbalance is important for model training—naive classifiers predicting all instances as ≤$50K would achieve 75.9% accuracy, making this a useful baseline for evaluating model performance.

Historical Context and Economic Interpretation

$50,000 in 1994 ≈ $103,000 in 2024 (adjusted for inflation). The income threshold represents upper-middle-class earnings for that era, roughly equivalent to the top 25% of individual earners.

The data captures the American workforce during a period of economic expansion following the early 1990s recession. Understanding this historical context is crucial when interpreting patterns—gender wage gaps, occupational distributions, and educational attainment rates have shifted significantly in the three decades since this data was collected.

Education and Income Correlation

Education level demonstrates the strongest correlation with income in this dataset. Advanced degrees dramatically increase the probability of earning above the $50K threshold.

View Source

SQL

SELECT education, COUNT(*) AS total, ROUND( 100.0 * SUM( CASE WHEN income = '>50K' THEN 1 ELSE 0 END ) / COUNT(*), 1 ) AS high_income_pct FROM adult_income.csv GROUP BY education ORDER BY high_income_pct DESC

Data

Education	High Income %
Doctorate	74.1
Prof-school	73.4
Masters	55.7
Bachelors	41.5
Assoc-voc	26.1
Assoc-acdm	24.8
Some-college	19
HS-grad	16
11th	5.6
Preschool	0
10 rows

Individuals with a Doctorate have a 74.1% probability of earning >$50K, compared to just 16% for high school graduates—a 4.6x difference highlighting the economic returns to education in 1994.

View Source

SQL

SELECT education, ROUND( 100.0 * SUM( CASE WHEN income = '>50K' THEN 1 ELSE 0 END ) / COUNT(*), 1 ) AS high_income_pct FROM adult_income.csv WHERE education IN ('Doctorate', 'HS-grad') GROUP BY education

Data

Education	High Income Pct
Doctorate	74.1
HS-grad	16
2 rows

Understanding the fnlwgt (Final Weight) Field

The fnlwgt column is one of the most misunderstood features in this dataset. It represents the Census Bureau's estimate of how many people in the US population each record represents, calculated using demographic stratification.

The weighting methodology groups people by state, age, race, and sex, then assigns weights so that the sample totals match known population controls. For example, if the sample under-represents young Hispanic females in California, records matching that profile receive higher weights.

For most ML tasks, exclude fnlwgt from your features. It's a sampling artifact, not a predictive attribute. However, it's valuable for weighted statistical analysis when you need population-level estimates rather than sample-level results.

Occupation Analysis

View Source

SQL

SELECT occupation, COUNT(*) AS total, ROUND( 100.0 * SUM( CASE WHEN income = '>50K' THEN 1 ELSE 0 END ) / COUNT(*), 1 ) AS high_income_pct FROM adult_income.csv WHERE occupation != '?' GROUP BY occupation ORDER BY high_income_pct DESC LIMIT 10

Data

Occupation	Total Workers	High Income %
Exec-managerial	4,066	48.4
Prof-specialty	4,140	44.9
Protective-serv	649	32.5
Tech-support	928	30.5
Sales	3,650	26.9
Craft-repair	4,099	22.7
Transport-moving	1,597	20
Adm-clerical	3,770	13.4
Machine-op-inspct	2,002	13.2
Farming-fishing	994	10.5
10 rows

Executive-managerial roles lead with a 48.4% high-income rate, followed by Professional-specialty at 44.9%. Service and manual labor occupations show rates below 15%.

View Source

SQL

SELECT occupation, ROUND( 100.0 * SUM( CASE WHEN income = '>50K' THEN 1 ELSE 0 END ) / COUNT(*), 1 ) AS high_income_pct FROM adult_income.csv WHERE occupation IN ('Exec-managerial', 'Prof-specialty') GROUP BY occupation ORDER BY high_income_pct DESC

Data

Occupation	High Income Pct
Exec-managerial	48.4
Prof-specialty	44.9
2 rows

Age and Income Patterns

Ages range from 17 to 90 years, with a mean of 38.6 years and median of 37 years.

View Source

SQL

SELECT MIN(age) AS min_age, MAX(age) AS max_age, ROUND(AVG(age), 1) AS avg_age FROM adult_income.csv

Data

Min Age	Max Age	Avg Age
17	90	38.6
1 row

View Source

SQL

SELECT CASE WHEN age < 25 THEN '17-24' WHEN age < 35 THEN '25-34' WHEN age < 45 THEN '35-44' WHEN age < 55 THEN '45-54' WHEN age < 65 THEN '55-64' ELSE '65+' END AS age_group, COUNT(*) AS total, SUM( CASE WHEN income = '>50K' THEN 1 ELSE 0 END ) AS high_income, ROUND( 100.0 * SUM( CASE WHEN income = '>50K' THEN 1 ELSE 0 END ) / COUNT(*), 1 ) AS high_income_pct FROM adult_income.csv GROUP BY age_group ORDER BY age_group

Data

Age Group	Total	High Income	High Income %
17-24	5,570	61	1.1
25-34	8,479	1,427	16.8
35-44	8,151	2,703	33.2
45-54	5,853	2,348	40.1
55-64	3,172	1,026	32.3
65+	1,336	276	20.7
6 rows

Peak earning years occur between 45-54, where the high-income rate reaches 40.1%. Young workers (17-24) have only a 1.1% probability of earning >$50K, reflecting entry-level positions and limited experience. The decline after 55 likely reflects early retirement and reduced work hours.

Gender Disparity Analysis

The dataset reveals significant gender disparity: 30.6% of males earn >$50K compared to only 10.9% of females—a 2.8x difference.

View Source

SQL

SELECT sex, COUNT(*) AS total, SUM( CASE WHEN income = '>50K' THEN 1 ELSE 0 END ) AS high_income, ROUND( 100.0 * SUM( CASE WHEN income = '>50K' THEN 1 ELSE 0 END ) / COUNT(*), 1 ) AS high_income_pct FROM adult_income.csv GROUP BY sex

Data

Sex	Total	High Income	High Income Pct
Male	21,790	6,662	30.6
Female	10,771	1,179	10.9
2 rows

View Source

SQL

SELECT sex, SUM( CASE WHEN income = '<=50K' THEN 1 ELSE 0 END ) AS low_income, SUM( CASE WHEN income = '>50K' THEN 1 ELSE 0 END ) AS high_income FROM adult_income.csv GROUP BY sex

Data

Gender	≤$50K	>$50K
Female	9,592	1,179
Male	15,128	6,662
2 rows

The dataset contains 21,790 males (66.9%) and 10,771 females (33.1%), reflecting 1994 labor force participation rates. This gender imbalance and income disparity make the dataset particularly valuable for algorithmic fairness research and bias auditing.

Working Hours and Capital Gains

High earners work an average of 45.5 hours per week compared to 38.8 hours for those earning ≤$50K—a 17% difference.

View Source

SQL

SELECT income, ROUND(AVG(hours_per_week), 1) AS avg_hours FROM adult_income.csv GROUP BY income

Data

Income	Avg Hours
>50K	45.5
<=50K	38.8
2 rows

Capital gains show dramatic differences: high earners average $4,006 in capital gains versus just $149 for lower earners—a 27x multiplier indicating wealth accumulation patterns.

View Source

SQL

SELECT income, ROUND(AVG(capital_gain), 2) AS avg_capital_gain FROM adult_income.csv GROUP BY income

Data

Income	Avg Capital Gain
>50K	4,006.14
<=50K	148.75
2 rows

Missing Values Analysis

Missing values are encoded as ' ?' (with a leading space) and appear in three columns:

| Column | Missing Count | Missing % | |--------|---------------|----------| | workclass | 1,836 | 5.6% | | occupation | 1,843 | 5.7% | | native_country | 583 | 1.8% |

Missing Value Strategies: - Remove rows: Drops ~7% of data (2,399 unique rows with any missing value) - Treat '?' as category: Preserves all data, may capture meaningful patterns (e.g., self-employed without formal occupation) - Mode imputation: Replace with most frequent value per column - Predictive imputation: Use other features to predict missing values - Use robust algorithms: XGBoost, LightGBM handle missing values natively

Baseline Model Performance

Typical accuracy scores on this dataset provide benchmarks for evaluating your models:

| Model | Accuracy | AUC-ROC | Notes | |-------|----------|---------|-------| | Naive Baseline (predict majority) | 75.9% | 0.50 | Always predicts ≤$50K | | Logistic Regression | ~84-85% | ~0.88 | Good interpretability | | Decision Tree | ~82-84% | ~0.85 | Prone to overfitting | | Random Forest | ~85-86% | ~0.90 | Strong baseline | | Gradient Boosting (XGBoost) | ~86-87% | ~0.92 | Current best practices | | Neural Network | ~85-86% | ~0.91 | Requires more tuning |

Note: Exact scores vary based on preprocessing choices, train/test splits, and hyperparameter tuning. The ~87% accuracy ceiling reflects inherent noise and the limited predictive power of available features.

Known Limitations and Ethical Considerations

When using this dataset, be aware of these important limitations:

- [ ] Temporal obsolescence: Data is 30 years old; labor markets, gender dynamics, and income distributions have changed significantly - [ ] Geographic limitation: US-only data; patterns don't generalize to other countries - [ ] Historical bias: Labels reflect 1994 societal biases in hiring, promotion, and compensation - [ ] Binary income threshold: Oversimplifies continuous income distribution - [ ] Self-reported data: Subject to reporting errors and social desirability bias - [ ] Protected attributes: Contains race and sex features that may lead to discriminatory models if used carelessly

For Fairness Research: This dataset is widely used to study algorithmic bias precisely because it exhibits disparate outcomes across protected groups. When building models, consider measuring equalized odds, demographic parity, and calibration across subgroups.

Sample Data Preview

Sample Records from the Dataset

#	Age	Workclass	Education	Occupation	Sex	Hours/Week	Income
1	39	State-gov	Bachelors	Adm-clerical	Male	40	<=50K
2	50	Self-emp-not-inc	Bachelors	Exec-managerial	Male	13	<=50K
3	38	Private	HS-grad	Handlers-cleaners	Male	40	<=50K
4	53	Private	11th	Handlers-cleaners	Male	40	<=50K
5	28	Private	Bachelors	Prof-specialty	Female	40	<=50K
6	37	Private	Masters	Exec-managerial	Female	40	<=50K
7	49	Private	9th	Other-service	Female	16	<=50K
8	52	Self-emp-not-inc	HS-grad	Exec-managerial	Male	45	>50K
8 rows

View Source

SQL

SELECT age, workclass, education, occupation, sex, hours_per_week, income FROM adult_income.csv LIMIT 8

Data

Age	Workclass	Education	Occupation	Sex	Hours/Week	Income
39	State-gov	Bachelors	Adm-clerical	Male	40	<=50K
50	Self-emp-not-inc	Bachelors	Exec-managerial	Male	13	<=50K
38	Private	HS-grad	Handlers-cleaners	Male	40	<=50K
53	Private	11th	Handlers-cleaners	Male	40	<=50K
28	Private	Bachelors	Prof-specialty	Female	40	<=50K
37	Private	Masters	Exec-managerial	Female	40	<=50K
49	Private	9th	Other-service	Female	16	<=50K
52	Self-emp-not-inc	HS-grad	Exec-managerial	Male	45	>50K
8 rows

Categorical Value Reference

Workclass Categories Explained

| Value | Description | |-------|-------------| | Private | For-profit private sector employment | | Self-emp-not-inc | Self-employed, business not incorporated | | Self-emp-inc | Self-employed, incorporated business | | Federal-gov | US federal government employee | | Local-gov | Local government employee | | State-gov | State government employee | | Without-pay | Unpaid family worker | | Never-worked | Never held a job |

Education Levels and Numeric Mapping

| Education | education_num | Description | |-----------|---------------|-------------| | Preschool | 1 | No formal education | | 1st-4th | 2 | Elementary (grades 1-4) | | 5th-6th | 3 | Elementary (grades 5-6) | | 7th-8th | 4 | Middle school | | 9th | 5 | High school freshman | | 10th | 6 | High school sophomore | | 11th | 7 | High school junior | | 12th | 8 | High school senior (no diploma) | | HS-grad | 9 | High school graduate | | Some-college | 10 | College, no degree | | Assoc-voc | 11 | Associate's (vocational) | | Assoc-acdm | 12 | Associate's (academic) | | Bachelors | 13 | Bachelor's degree | | Masters | 14 | Master's degree | | Prof-school | 15 | Professional school (MD, JD) | | Doctorate | 16 | Doctoral degree (PhD) |

Occupation Categories Explained

| Value | Description | |-------|-------------| | Exec-managerial | Executive and managerial positions | | Prof-specialty | Professional specialty (doctors, lawyers, engineers) | | Tech-support | Technical support roles | | Sales | Sales occupations | | Adm-clerical | Administrative and clerical | | Protective-serv | Police, firefighters, security | | Priv-house-serv | Private household service | | Handlers-cleaners | Material handlers, cleaners | | Machine-op-inspct | Machine operators, inspectors | | Transport-moving | Transportation and material moving | | Craft-repair | Skilled trades and repair | | Farming-fishing | Agriculture and fishing | | Armed-Forces | Military personnel | | Other-service | Other service occupations |

Table Overview

adult_income

Contains 32,561 rows and 15 columns. Column types: 6 numeric, 9 text.

32,561 rows15 columns

adult_income

32,561

rows

columns

Data Preview

Scroll to see more

age	workclass	fnlwgt	education	education_num	marital_status	occupation	relationship	race	sex	capital_gain	hours_per_week	native_country	income
39	State-gov	77,516	Bachelors	13	Never-married	Adm-clerical	Not-in-family	White	Male	2,174	40	United-States	<=50K
50	Self-emp-not-inc	83,311	Bachelors	13	Married-civ-spouse	Exec-managerial	Husband	White	Male	0	13	United-States	<=50K
38	Private	215,646	HS-grad	9	Divorced	Handlers-cleaners	Not-in-family	White	Male	0	40	United-States	<=50K
53	Private	234,721	11th	7	Married-civ-spouse	Handlers-cleaners	Husband	Black	Male	0	40	United-States	<=50K
28	Private	338,409	Bachelors	13	Married-civ-spouse	Prof-specialty	Wife	Black	Female	0	40	Cuba	<=50K

Row 1

age39

workclassState-gov

fnlwgt77,516

educationBachelors

education_num13

+10 more columns

Row 2

age50

workclassSelf-emp-not-inc

fnlwgt83,311

educationBachelors

education_num13

+10 more columns

Row 3

age38

workclassPrivate

fnlwgt215,646

educationHS-grad

education_num9

+10 more columns

Showing 5 of 32,561 rows

Data Profile

32,561

rows

columns

100%

complete

23.3 MB

estimated size

Column Types

6 Numeric9 Text

High-Cardinality Columns

Columns with many unique values (suitable for identifiers or categorical features)

fnlwgt(21,648 unique values)

Data Dictionary

adult_income

Column	Type	Description	Example
`age`	numeric	Age in years (28 to 53)	39, 50
`workclass`	string	Class category	"State-gov", "Self-emp-not-inc"
`fnlwgt`	numeric	Numeric value (range: 77,516 - 338,409)	77516, 83311
`education`	string	Education	"Bachelors", "Bachelors"
`education_num`	numeric	Numeric value (range: 7 - 13)	13, 13
`marital_status`	string	Status indicator	"Never-married", "Married-civ-spouse"
`occupation`	string	Occupation	"Adm-clerical", "Exec-managerial"
`relationship`	string	Relationship	"Not-in-family", "Husband"
`race`	string	Race	"White", "White"
`sex`	string	Gender	"Male", "Male"
`capital_gain`	numeric	Numeric value (range: 0 - 2,174)	2174, 0
`capital_loss`	numeric	Numeric value (range: 0 - 0)	0, 0
`hours_per_week`	numeric	Numeric value (range: 13 - 40)	40, 13
`native_country`	string	Count or quantity	"United-States", "United-States"
`income`	string	Income	"<=50K", "<=50K"