Software Defect Prediction Dataset: 284,676 Code File Instances from Five Open-Source Repositories

A benchmark dataset for machine-learning-driven software defect prediction, comprising 284,676 code file observations from five open-source repositories: Elasticsearch (165,032 instances, 21.2% defect rate), Spring Boot (56,425, 11.0%), Hadoop (49,047, 17.0%), Kafka (13,158, 37.6%), and Express (1,014, 37.5%). Overall defect rate: 19.3%.

 

Each observation records seven process metrics from git history: commit count, unique developer count, lines added, lines deleted, code churn, file age, and commit frequency — plus a binary defect label. A trained Gradient Boosting classifier (AUC=0.899) and Random Forest (AUC=0.904, F1=0.682) are included as scikit-learn pipelines.

 

Note: The column bug_fix_commits is included for transparency but must be excluded from predictive models — it is a retrospective leakage feature. See DATA_README.md for details.

 

Code: https://github.com/javvadivijayprasad/ai-defect-prediction-repository-analytics