A standard sklearn LogisticRegression trained on the UCI Adult Income dataset, predicting whether annual income exceeds $50,000 USD based on US Census features.
This model is hosted publicly as a reference target for fairness testing demonstrations, particularly with the verosynthea-validator. It is not intended for production use.
Training Data
UCI Adult Income dataset (1994 US Census extract). ~48,000 rows, 14 features. Loaded via the scikit-learn/adult-census-income HuggingFace dataset.
Features
Feature
Type
Description
age
numeric
Age in years
workclass
categorical
Employment type (Private, Self-emp, Gov, ...)
fnlwgt
numeric
Census sampling weight
education
categorical
Highest education level
education-num
numeric
Education as ordinal number
marital-status
categorical
Marital status
occupation
categorical
Occupation category
relationship
categorical
Relationship to householder
race
categorical
Race
sex
categorical
Sex
capital-gain
numeric
Capital gains
capital-loss
numeric
Capital losses
hours-per-week
numeric
Hours worked per week
native-country
categorical
Country of birth
Performance
Accuracy on UCI Adult test set: ~84%
Reproduces standard published baselines for this dataset
Baseline comparator for fairness mitigation experiments
Out-of-Scope Use
DO NOT use this model for any real income classification task. The training data is from 1994 US Census and is unsuitable for current or non-US populations.
Running this model on Australian populations (e.g., via AUSynth data) produces large demographic bias gaps. This is by design — it demonstrates why models trained on one population should be tested before deployment in another.
