Fine-tuning Pre-trained Language Models for Software Traceability: A Comprehensive Empirical Study

Source code for the TOSEM paper Fine-tuning Pre-trained Language Models for Software Traceability: A Comprehensive Empirical Study.

Folder

• data is used to store the processed datasets.
• distill folder contains the data format used in the distillation step dataset.py, the configuration file tiny_bert_config.json for student model and the distillation file bertdistill.py. Run the file run.sh for KD.
• models contains training/testing files and trained models. Run the file run.sh for Full-FT/Partial-FT/PEFT.

Environment

• python 3.9.13
• matplotlib 3.9.4
• nltk 3.9.1
• numpy 1.24.3
• pandas 1.3.4
• peft 0.15.2
• tokenizers 0.21.1
• torch 1.13.0+cu116
• torchaudio 0.13.0+cu116
• torchvision 0.14.0+cu116
• transformers 4.47.0

Datasets

This study employs six large-scale open-source project datasets (Ambari, Calcite, Groovy, Ignite, Isis, Netbeans), originally constructed and used by EALink, a state-of-the-art method for issue–commit link recovery.

How to run

1.Download the dataset to data folder & Download six models from HuggingFace

2.Run distill/run.sh and models/run.sh