Replication Package for Synthesizing Static Rules for Code Smell Detection

A 2026 replication package by Zijie Huang for research on automated code smell detection. It includes the MLCQ benchmark with 14,739 annotations from 522 repositories, results from 40 developer interviews, and evaluation data from 1,840 developer assessments. The package contains datasets, source code for a Java subsystem, and Python scripts for reproducing machine learning experiments.

Use Cases

• Benchmarking code smell detection models based on the provided MLCQ dataset with 14,739 annotations.
• Training or evaluating machine learning models like CodeBERT-fusion or threshold optimization methods described in the package.
• Conducting developer studies on code quality perceptions using the provided 1,840 evaluation instances and 40 interview transcripts.
• Reproducing statistical analyses for research on rule-based vs. ML-based detection methods.

Strengths

• Includes a substantial benchmark dataset (MLCQ) with 14,739 code smell annotations.
• Contains human evaluation data from 1,840 developer assessments and 40 qualitative interviews.
• Provides full replication materials including source code (~5.3 KLOC), model implementations, and pre-computed results.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row counts for some constituent datasets are unknown, which may limit suitability assessment.
• The 917.8 KB size suggests a limited scope, primarily containing aggregated results and scripts rather than raw, large-scale source code corpora.

Provenance

Source

figshare, author Zijie Huang.

Collection Method

Likely compiled from open-source repositories (e.g., MLCQ benchmark) and original research experiments.

Freshness

Last updated 2026-04-17 05:58:55.