TRD: Test Refactoring Dataset

TRD is a reproducible benchmark for predicting the refactoring of Java test methods. It pairs a benchmark dataset with the evaluation code and the deterministic split protocol needed to reproduce the results reported in the accompanying paper. The benchmark defines two tasks: - Track A (binary): predict whether a test method will be refactored. - Track B (multi-label): predict which refactoring types apply to a test method. Central finding: models that look strong under a project-mixed split collapse toward the prevalence baseline under a held-out-project (cross-project) split, and full fine-tuning does not close the gap. TRD is intended as a target for methods that aim to generalize refactoring prediction to unseen projects. Contents: the Track A dataset (122,098 samples, 44 projects) with engineered tabular features, the Track B dataset (46,713 samples, 40 projects, 21 modeled refactoring types), the evaluation and figure-generation code, experiment configurations with seeds, precomputed per-fold metrics and out-of-fold predictions, and provenance with checksums. Licensing: the benchmark annotations, feature tables, dataset structure, and split protocol are released under CC BY 4.0. The code is released under the MIT License (see LICENSE-CODE.txt). The "method_code" field contains source code extracted from public open-source Java projects; those snippets remain under the licenses of their respective upstream projects and are included for research and reproducibility. The CC BY license covers only the original annotations, feature engineering, and dataset structure, not the upstream source code. Each sample records its project and commit so the upstream provenance can be traced.