ROSE: A Transformer-based Recommender for Refactoring Operations from Architectural Smell

This dataset accompanies the study ROSE: A Transformer-based Recommender for Refactoring Operations from Architectural Smells. It contains the full replication package for training and evaluating transformer models (CodeBERT and CodeT5) on the task of recommending refactorings to remove architectural smells. The dataset is class-balanced, ensuring equal representation of the three architectural smells and their associated refactorings. The balancing was achieved through random undersampling of over-represented classes, ensuring fair evaluation and comparability across models. The corpus includes: Labeled code snippets (Java) with associated refactoring types Refactoring labels mapped to architectural smells: God Class → Extract Method Cyclic Dependency → Move Class Hub-like Dependency → Pull Up Method