Evaluating Embedding Representations for Multiclass Code Smell Detection: A Comparative Study of CodeBERT and General-Purpose Embeddings

This study uses a dataset derived from the Crowdsmelling dataset, containing instances of Long Method, God Class, and Feature Envy extracted from open-source Java projects. The original instances were consolidated into a single multiclass dataset and manually mapped to their corresponding source code fragments. As part of preprocessing, line comments and block comments were removed, while line breaks were preserved to maintain the structural organization of the code.