Health Equity Benchmark (HEB): A Synthetic Dataset for Responsible AI Education and Fairness Evaluation

Responsible AI education needs datasets and workflows that let learners observe fairness failures, quantify them, and test mitigation strategies without handling sensitive patient records. This need is especially acute in African and other resource constrained teaching environments, where access to licensed health data, specialized infrastructure, and reproducible instructional materials is often limited. We present the Health Equity Benchmark (HEB), a synthetic healthcare dataset and accompanying browser based tutorial for fairness oriented machine learning education. HEB contains 16{,}000 synthetic patient records with demographic, socioeconomic, access to care, lifestyle, and clinical variables plus a binary outcome label. The tutorial supports exploratory analysis, baseline modeling, fairness evaluation with demographic parity difference and equalized odds difference, post processing mitigation with Fairlearn, and fairness oriented neural network training. Across Logistic Regression, Random Forest, XGBoost, and neural models, HEB produces interpretable fairness performance trade offs that are useful for classroom instruction. A usability study with 130 participants further suggests that the dataset and tutorial improve self reported understanding of fairness metrics and increase learner confidence. HEB is intended as open educational infrastructure for health focused Responsible AI teaching, especially in contexts where privacy preserving, easy to run, and African relevant learning resources are urgently needed.