torch 1.7.0
transformers 4.18.0
Download codebert-base and unixcoder-base, and move the files to ./microsoft/codebert-base and ./microsoft/unixcoder-base
The error classification task requires that we assign the same label to a <code, line> pair that have the same run-time error type. Models are evaluated by accuracy.
We use CodeNet Python800 dataset on this task.
The process for dataset is similar with clone detection.
python preprocess_cls.pyAfter preprocessing dataset, you can obtain the three .jsonl files,
i.e. train_exception_cls.jsonl, valid_exception_cls.jsonl, test_exception_cls.jsonl.
The processed .jsonl files are also at ./dataset.
For each file, each line in the uncompressed file represents one function. One row is illustrated below.
- code: the path to the source code, for example,
Python800/Project_CodeNet_Python800/p00000/s003971419.py - label: the error type of the <code, line> pair
- index: the index of example
- line: the line of code that triggers the error
You can find the raw and processed dataset shared on Google Drive.
Please cite our work in your publications if it helps your research: ToDo