GenAI Systems Engineer

CASE STUDY

Software Defect Prediction

An Empirical Investigation into Within-Project Software Defect Prediction using CodeBERT and Static Code Metrics.

Abstract

This thesis presents CodeBERT-SMCH, a hybrid Transformer architecture that combines CodeBERT semantic embeddings with 20 static code metrics for within-project software defect prediction. The model outperformed reference LLM baselines, achieving a Recall of 0.81 and an F1-score of 0.62. Implemented using PyTorch, Hugging Face Transformers, and Scikit-learn on NVIDIA T4 GPUs.

Key Results

0.81Recall
0.62F1 Score