JetBrains-Research/lca-project-level-code-completion

🏟️ Long Code Arena (Project-level code completion)

This is the benchmark for Project-level code completion task as part of the 🏟️ Long Code Arena benchmark. Each datapoint contains the file for completion, a list of lines to complete with their categories (see the categorization below), and a repository snapshot that can be used to build the context. 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.

How-to

Load the data via load_dataset:

from datasets import load_dataset

config_names = [
  'small_context',
  'medium_context',
  'large_context',
  'huge_context'
]

ds = load_dataset('JetBrains-Research/lca-project-level-code-completion', config_name, split='test')

Dataset Structure

Datapoints in the dataset have the following structure:

• repo – repository name in the format {GitHub_user_name}__{repository_name}
• commit_hash – commit hash of the repository
• completion_file – dictionary with the completion file content in the following format:
  • filename – path to the completion file
  • content – content of the completion file
• completion_lines – dictionary where the keys are categories of lines and values are a list of integers (numbers of lines to complete). The categories are:
  • committed – line contains at least one function or class from the files that were added on the completion file commit
  • inproject – line contains at least one function or class from the repository snapshot at the moment of completion
  • infile – line contains at least one function or class from the completion file
  • common – line contains at least one function or class with common names, e.g., main, get, etc.
  • non_informative – line that was classified to be non-informative, e.g., too short, contains comments, etc.
  • random – other lines.
• repo_snapshot – dictionary with a snapshot of the repository before the commit. It has the same structure as completion_file, but filenames and contents are orginized as lists.
• completion_lines_raw – same as completion_lines, but before sampling

How we collected the data

To collect the data, we cloned repositories from GitHub where the main language is Python. The completion file for each datapoint is a .py file that was added to the repository in a commit. The state of the repository before this commit is the repo snapshot.

The dataset configurations are based on the number of characters in .py files from the repository snapshot:

• small_context – less than 48K characters;
• medium_context – from 48K to 192K characters;
• large_context – from 192K to 768K characters;
• huge_context – more than 768K characters.

Datasets Stats

Dataset	Number of datapoints	Number of repositories	Number of commits
small_context	144	46	63
medium_context	224	80	175
large_context	270	75	219
huge_context	296	75	252

Completion File

Dataset	Completion file lines, min	Completion file lines, max	Completion file lines, median
small_context	201	1916	310.5
medium_context	200	1648	310.0
large_context	200	1694	278.0
huge_context	200	1877	313.5

Repository Snapshot .py files

Dataset	Context py files number, min	Context py files number, max	Context py files number, median	Context py lines, median
small_context	0	52	4.0	128.0
medium_context	3	117	34.0	3786.0
large_context	3	255	84.0	15466.5
huge_context	47	5227	261.0	49811.0

Repository Snapshot non .py files

Dataset	Context non-py files number, min	Context non-py files number, max	Context non-py files number, median	Context non-py lines, median
small_context	1	1044	19.5	1227.0
medium_context	3	3977	64.5	9735.0
large_context	8	2174	155.0	18759.0
huge_context	24	7687	262.0	60163.0

Line Counts:

Dataset	infile	inproject	common	commited	non-informative	random	all
small_context	1430	95	500	1426	532	703	4686
medium_context	2224	2236	779	1495	858	1084	8676
large_context	2691	2595	693	1322	1019	1311	9631
huge_context	2608	2901	692	1019	1164	1426	9810

Scores

You can find the results of running various models on this dataset in our leaderboard.

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.