JetBrains-Research/lca-commit-message-generation

🏟️ Long Code Arena (Commit message generation)

This is the benchmark for the Commit message generation task as part of the 🏟️ Long Code Arena benchmark.

The dataset is a manually curated subset of the Python test set from the 🤗 CommitChronicle dataset, tailored for larger commits.

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

How-to

from datasets import load_dataset

dataset = load_dataset("JetBrains-Research/lca-cmg", split="test")

Note that all the data we have is considered to be in the test split.

Note. Working with git repositories under repos directory is not supported via 🤗 Datasets. See Git Repositories section for more details.

About

Overview

In total, there are 163 commits from 34 repositories. For length statistics, refer to the notebook in our repository.

Dataset Structure

The dataset contains two kinds of data: data about each commit (under commitchronicle-py-long folder) and compressed git repositories (under repos folder).

Commits

Each example has the following fields:

Field	Description
repo	Commit repository.
hash	Commit hash.
date	Commit date.
license	Commit repository's license.
message	Commit message.
mods	List of file modifications from a commit.

Each file modification has the following fields:

Field	Description
change_type	Type of change to current file. One of: ADD, COPY, RENAME, DELETE, MODIFY or UNKNOWN.
old_path	Path to file before change (might be empty).
new_path	Path to file after change (might be empty).
diff	git diff for current file.

Data point example:

{'hash': 'b76ed0db81b3123ede5dc5e5f1bddf36336f3722',
 'repo': 'apache/libcloud',
 'date': '05.03.2022 17:52:34',
 'license': 'Apache License 2.0',
 'message': 'Add tests which verify that all OpenStack driver can be instantiated\nwith all the supported auth versions.\nNOTE: Those tests will fail right now due to the regressions being\nintroduced recently which breaks auth for some versions.',
 'mods': [{'change_type': 'MODIFY',
    'new_path': 'libcloud/test/compute/test_openstack.py',
    'old_path': 'libcloud/test/compute/test_openstack.py',
    'diff': '@@ -39,6 +39,7 @@ from libcloud.utils.py3 import u\n<...>'}],
}    

Git Repositories

The compressed Git repositories for all the commits in this benchmark are stored under repos directory.

Working with git repositories under repos directory is not supported directly via 🤗 Datasets. You can use huggingface_hub package to download the repositories. The sample code is provided below:

import tarfile
from huggingface_hub import list_repo_tree, hf_hub_download


data_dir = "..."  # replace with a path to where you want to store repositories locally

for repo_file in list_repo_tree("JetBrains-Research/lca-commit-message-generation", "repos", repo_type="dataset"):
    file_path = hf_hub_download(
        repo_id="JetBrains-Research/lca-commit-message-generation",
        filename=repo_file.path,
        repo_type="dataset",
        local_dir=data_dir,
    )

    with tarfile.open(file_path, "r:gz") as tar:
        tar.extractall(path=os.path.join(data_dir, "extracted_repos"))

For convenience, we also provide a full list of files in paths.json.

After you download and extract the repositories, you can work with each repository either via Git or via Python libraries like GitPython or PyDriller.

Extra: longer context

Full Files

To facilitate further research, we additionally provide full contents of modified files before and after each commit in full_files dataset config. full split provides the whole files, and the remaining splits truncates each file given the maximum allowed number of tokens n. The files are truncated uniformly, essentially, limiting the number of tokens for each file to max_num_tokens // num_files. We use DeepSeek-V3 tokenizer to obtain the number of tokens.

from datasets import load_dataset

dataset = load_dataset("JetBrains-Research/lca-commit-message-generation", 
                       "full_files",
                       split="16k"  # should be one of: '4k', '8k', '16k', '32k', '64k', 'full'
                       )

Each example has the following fields:

• repo: commit repository
• hash: commit hash
• mods: commit modification (combined into a single diff)
• files: a list of dictionaries, where each corresponds to a specific file changed in the commit and has the following keys:
  • old_path: file path before the commit
  • old_contents: file contents before the commit
  • new_path: file path after the commit
  • old_contents: file contents after the commit

Retrieval

To facilitate further research, we additionally provide context for each commit as retrieved by BM25 retriever in retrieval_bm25 dataset config. For each commit, we run BM25 over all .py files in the corresponding repository at the state before the commit (excluding the files that were changed in this commit). We retrieve up to 50 files most relevant to the commit diff, and then, given the maximum allowed number of tokens n, we add files until the total context length (including diff) in tokens returned by the DeepSeek-V3 tokenizer exceeds n, possibly trunctating the last included file.

To access these, run the following:

from datasets import load_dataset

dataset = load_dataset("JetBrains-Research/lca-commit-message-generation", 
                       "retrieval_bm25",
                       split="16k"  # should be one of: '4k', '8k', '16k', '32k', '64k'
                       )

Each example has the following fields:

• repo: commit repository
• hash: commit hash
• mods: commit modification (combined into a single diff)
• context: context retrieved for the current commit; a list of dictionaries, where each corresponds to a specific file and has the following keys:
  • source: file path
  • content: file content

🏷️ Extra: commit labels

To facilitate further research, we additionally provide the manual labels for all the 858 commits that made it through initial filtering. The final version of the dataset described above consists of commits labeled either 4 or 5.

How-to

from datasets import load_dataset

dataset = load_dataset("JetBrains-Research/lca-commit-message-generation", "labels", split="test")

Note that all the data we have is considered to be in the test split.

About

Dataset Structure

Each example has the following fields:

Field	Description
repo	Commit repository.
hash	Commit hash.
date	Commit date.
license	Commit repository's license.
message	Commit message.
label	Label of the current commit as a target for CMG task.
comment	Comment for a label for the current commit (optional, might be empty).

Labels are in 1–5 scale, where:

• 1 – strong no
• 2 – weak no
• 3 – unsure
• 4 – weak yes
• 5 – strong yes

Data point example:

{'hash': '1559a4c686ddc2947fc3606e1c4279062cc9480f',
 'repo': 'appscale/gts',
 'date': '15.07.2018 21:00:39',
 'license': 'Apache License 2.0',
 'message': 'Add auto_id_policy and logs_path flags\n\nThese changes were introduced in the 1.7.5 SDK.',
 'label': 1,
 'comment': 'no way to know the version'}

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.