CAMS Replication: Code Smell Detection with Inconsistent Annotations

A 64.1 MB replication package for research on code smell detection, last updated on 2026-05-28. It includes raw and processed datasets, pre-computed code embeddings from models like CodeBERT and UniXcoder, and scripts to reproduce experimental results. The package is authored anonymously and shared under a CC-BY-4.0 license on figshare.

Use Cases

• Benchmarking code smell detection models based on the provided MLCQ-derived datasets with manual labels and software metrics.
• Reproducing experiments on leveraging inconsistent annotations for training, as described in the accompanying research paper.
• Utilizing pre-computed code embeddings from large models to reduce computational overhead for model training or evaluation.
• Conducting inter-rater agreement analysis on code smell annotations based on the provided statistical scripts.

Strengths

• Includes a complete replication package with source code, raw data, and automated scripts for full reproducibility.
• Provides pre-computed embedding vectors from multiple large models (CodeBERT, CodeT5p, Jina, UniXcoder) to reduce GPU resource requirements.
• Repository structure is designed to align with the logical narrative of the research paper for easier navigation.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment for large-scale training.
• The dataset's foundational source, MLCQ, may reflect specific biases in code selection and annotation.

Provenance

Source

Derived from the MLCQ dataset, as indicated in the description.

Collection Method

Likely involves manual labeling of code samples and extraction of software metrics, with subsequent processing for machine learning.

Freshness

Last updated 2026-05-28 15:05:19; freshness should be verified.