MLCQ‑Experiments Dataset Overview

Source

• MLCQ Dataset: MLCQ dataset

Purpose

• Supports the paper "Exploring NLP Techniques for Code Smell Detection: A Comparative Study".
• Enables comparison of NLP‑based models with baseline approaches for detecting code smells.

Processing Steps

1. Environment Setup:
   • Create a conda environment and install dependencies.
   • Commands:

   conda create -n mlcqenv python=3.10
   conda activate mlcqenv
   conda install -f requirements.txt
   

2. Data Extraction:
   • Set a GitHub token for API access.
   • Export token: export GITHUB_TOKEN=<your_github_token>
   • Run the extractor: python DataExtractor.py
3. Baseline Model:
   • Use the J48 decision‑tree as baseline.
   • Compute code metrics with the Designite tool.
   • Steps:
     1. python baseline/MetricsExtractor.py (prepare .java files)
     2. python baseline/DesigniteRun.py (run Designite)
     3. python baseline/DatasetCreator.py (final dataset)
     4. python train.py (train & test)
4. Model Training:
   • Train BiLSTM with Attention:

   python bilstm_attn_train.py --batch_size 16 --epochs 20 --learning_rate 0.0001 --hidden_dim 512 --num_layers 2
   

   • Train CodeBERT:

   python bert.py
   

Dependencies

• Designite Tool: Designite tool
• CodeBERT Pre‑trained Model: CodeBERT on Hugging Face

Authors

• Djamel Mesbah
• Nour El Madhoun
• Hani Chalouati
• Khaldoun Al Agha