Function-Level Code Changes for Defect Prediction in C/C++ Projects

13,432 function-level modifications from 22 large-scale C/C++ projects, curated for software defect prediction research. The dataset includes 3,164 defective and 10,268 clean modifications. It was created by Doha Nam for the FSE 2026 paper 'ReDef' and released in April 2026.

Use Cases

• Train and evaluate code language models to classify function-level modifications as defective or clean.
• Analyze the relationship between specific code change features and the introduction of defects across 22 C/C++ projects.
• Benchmark the performance of different models on the task of just-in-time software defect prediction.
• Study the characteristics of defective versus clean code changes to inform software quality assurance practices.

Strengths

• Contains 13,432 labeled function-level modifications (3,164 defective, 10,268 clean).
• Curated from 22 large-scale, real-world C/C++ projects.
• Includes a complete replication package with scripts for dataset collection and model evaluation.

Limitations

• Dataset size is relatively small (13.4k samples) for training large language models.
• Limited to C and C++ programming languages, not generalizable to other languages.
• Class imbalance exists with approximately 4 times more clean samples than defective ones.

Provenance

Source

22 large-scale open-source C/C++ projects.

Collection Method

Curated function-level modifications from software repositories, labeled as defective or clean.

Freshness

Released in April 2026 for a paper accepted at FSE 2026.