JetBrains-Research/lca-bug-localization

🏟️ Long Code Arena (Bug localization)

This is the benchmark for the Bug localization task as part of the 🏟️ Long Code Arena benchmark.

The bug localization problem can be formulated as follows: given an issue with a bug description and a repository snapshot in a state where the bug is reproducible, identify the files within the repository that need to be modified to address the reported bug.

The dataset provides all the required components for evaluation of bug localization approaches in real project-level large-scale data collected from GitHub, including:

• Bug issue description;
• Repositories, from which the content can be extracted at the state of the commit SHA where the bug is reproducible;
• List of files that should be changed in order to solve the bug;
• Other additional data and metrics that can be useful in developing new approaches.

All the repositories are published under permissive licenses (MIT, Apache-2.0, BSD-3-Clause, and BSD-2-Clause). The datapoints can be removed upon request.

The collected dataset was carefully filtered, enhanced with useful metrics and, what's more, manually labeled, which assures the data quality and provides a golden subset of good examples for evaluation.
Moreover, the dataset was split into several categories, namely:

Category	Description	Number of data points
py	Only .py files in changes	4,339
java	Only .java files in changes	2,522
kt	Only .kt files in changes	618

...and splits, namely:

Split	Description
dev	All collected datapoints
test	Manually verified datapoints
train	Rest of the datapoint from dev without test

The results of evaluation of various bug localization approaches can be found in the Long Code Arena 🏟 leaderboard.

The following sections describe the utilities around the dataset, as well as dataset content.

How-to

• Load the data via load_dataset:

  from datasets import load_dataset
  
  # Select a configuration from ["py", "java", "kt"]
  configuration = "py"
  # Select a split from ["dev", "train", "test"]
  split = "dev"
  # Load data
  dataset = load_dataset("JetBrains-Research/lca-bug-localization", configuration, split=split)  
  

• Load repos via hf_hub_download:

  from huggingface_hub import hf_hub_download
  from datasets import load_dataset
  import zipfile
  import os
  
  # Load json with list of repos' .tar.gz file paths
  paths_json = load_dataset("JetBrains-Research/lca-bug-localization", data_files="repos.json")
  
  # Load each repo in .tar.gz format, unzip, delete archive
  repos = paths_json[configuration][0]
  repos_path = "path/to/repos"
  
  for i, repo_zip_path in enumerate(repos):
      repo_name = os.path.basename(repo_zip_path).split('.')[0]
      repo_path = os.path.join(repos_path, repo_name)
  
      # Load repo zip
      local_repo_zip_path = hf_hub_download(
          "JetBrains-Research/lca-bug-localization",
          filename=repo_zip_path,
          repo_type="dataset",
          local_dir="path/to/zips"
      )
  
      # Unzip repo
      with zipfile.ZipFile(local_repo_zip_path, 'r') as zip_ref:
          zip_ref.extractall(repo_path)
      os.remove(local_repo_zip_path)
  

• Data streaming via HFDataSource. Besides data loading, HFDataSource returns a datapoint for running the baseline along with the content of the repository at the state where the bug is reproducible (aka base_sha commit of the pull request that resloves the bug issue).
  All source code for working with the Git history of repositories (commits, diffs, etc.) is available in git_utils.py, as is an example of baselines, utilizing this dataset.

Bug localization data

Each datapoint contains the main fields, as well as additional metrics calculated on them.
The main fields are:

Field	Description
repo_owner	Owner of the repository with the bug issue.
repo_name	Name of the repository with the bug issue.
issue_url	GitHub link to the issue https://github.com/{repo_owner}/{repo_name}/issues/{issue_id}.
pull_url	GitHub link to the pull request https://github.com/{repo_owner}/{repo_name}/pull/{pull_id}.
comment_url	GitHub link to the comment with a reference from pull request to issue https://github.com/{repo_owner}/{repo_name}/pull/{pull_id}#issuecomment-{comment_id}.
issue_title	Issue title.
issue_body	Issue body.
base_sha	Base SHA of the pull request.
head_sha	Head SHA of the pull request.
diff_url	Link to the diff between the base and the head SHA https://github.com/{repo_owner}/{repo_name}/compare/{base_sha}...{head_sha}.
diff	Content of the diff.
pull_create_at	Date of pull request creation in the yyyy-mm-ddThh:mm:ssZ format.
repo_stars	Number of stars of the repo.
changed_files*	List of the changed files parsed from diff.
repo_language	Main programming language used in the repository.
repo_languages	All programming languages used in the repository.
repo_license	License assigned to the repository.

* Excluding test files that do not contain bug causes, rather changed in order to add tests for proving that the bug is gone.

The metrics-related fields are:

Field	Description
changed_files_exts	Dictionary from the extension of changed files to their count.
changed_files_count	Number of changed files.
java_changed_files_count	Number of changed .java files.
kt_changed_files_count	Number of changed .kt files.
py_changed_files_count	Number of changed .py files.
code_changed_files_count	Number of changed .java, .kt, or .py files.
repo_symbols_count*	Number of symbols in the repository.
repo_tokens_count*	Number of tokens** in the files of the repository.
repo_lines_count*	Number of lines in the files of the repository.
repo_files_without_tests_count*	Number of files in the repository.
changed_symbols_count*	Number of symbols in the changed lines in diff.
changed_tokens_count*	Number of tokens** in the changed lines in diff.
changed_lines_count*	Number of changed lines in diff (including added and deleted).
changed_files_without_tests_count*	Number of files in diff.
issue_symbols_count*	Number of symbols in the issue body.
issue_words_count*	Number of words in the issue body (separated by space symbols).
issue_tokens_count*	Number of tokens** in the issue body.
issue_lines_count*	Number of text lines in the issue body (separated by \\n).
issue_links_count	Number of links ([...](...)) present in the issue body.
issue_code_blocks_count	Number of code blocks (```...```) present in the issue body.

* Excluding test files that do not contain bug causes, rather changed in order to add tests for proving that the bug is gone.
** Using GPT-4 tokenizer via ticktoken.

Repositories data

The compressed repositories are provided in the repos section, separately from datapoints, to provide access to the various stages of repositories, saving their initial structure, as well as to reuse their content for different datapoints connected to the same repository. To extract the required information from repositories, each zip file should be unarchived. Afterwords, we recommend to use GitPython or PyDriller Python libraries to navigate through the repository history and extract its content on the required commit or calculate diff. Most of the required utility methods are provided in our repository in the git_utils.py file, so you may reuse them to access the required repository data.

Citing

@article{bogomolov2024long,
  title={Long Code Arena: a Set of Benchmarks for Long-Context Code Models},
  author={Bogomolov, Egor and Eliseeva, Aleksandra and Galimzyanov, Timur and Glukhov, Evgeniy and Shapkin, Anton and Tigina, Maria and Golubev, Yaroslav and Kovrigin, Alexander and van Deursen, Arie and Izadi, Maliheh and Bryksin, Timofey},
  journal={arXiv preprint arXiv:2406.11612},
  year={2024}
}

You can find the paper here.