README.md

metadata

configs:
  - config_name: repo_split_balanced
    data_files:
      - split: train
        path: repo_split_balanced/train-*
      - split: validation
        path: repo_split_balanced/validation-*
      - split: test
        path: repo_split_balanced/test-*
  - config_name: random_shuffle
    data_files:
      - split: train
        path: random_shuffle/train-*
      - split: validation
        path: random_shuffle/validation-*
      - split: test
        path: random_shuffle/test-*
task_categories:
  - text-retrieval
language:
  - en
  - zh

ArkTS-CodeSearch-Pro: A Cleaned and License-Aware ArkTS Dataset

Original Dataset | Pro Dataset | Paper | Evaluation Code | Dataset Processing Code

This dataset is an improved version of ArkTS-CodeSearch. It contains function-level information from ArkTS (.ets) projects, including function source code, docstrings, AST representations, and repository metadata. It is intended for code retrieval, code understanding, and AST-based research.

Compared with the original dataset, this Pro version provides the following improvements:

• License-aware collection: only repositories with locally identified open-source licenses are included. The supported license families include Apache-2.0, MIT, BSD, GPL, LGPL, and MPL.
• Repository quality filtering: only repositories with more than zero stars are retained.
• Comment normalization: comment delimiters such as //, /*, */, and leading * are removed from the docstrings.
• Noise removal: copyright/license headers, code-only comments, and cleaned docstrings shorter than five characters are excluded.
• Stronger deduplication: complete function texts and cleaned queries are globally unique.
• Repository-isolated evaluation: repo_split_balanced groups repositories by normalized nwo across GitHub, Gitee, and GitCode, preventing the same cross-platform repository name from appearing in multiple splits.
• Stable split sizes: both configurations are approximately 80%/10%/10%.

The final dataset contains 31,164 examples, all extracted from .ets files. The sha column was removed because commit SHA metadata was unavailable for most source repositories; function_sha is retained as a function-content identifier.

Dataset Structure

The dataset provides two configurations:

• repo_split_balanced: repository-level split with normalized nwo repository isolation. It contains 24,932 training, 3,116 validation, and 3,116 test examples.
• random_shuffle: globally shuffled split with seed 20260714. It contains 24,931 training, 3,117 validation, and 3,116 test examples.

Features / Columns

Field	Type	Description
nwo	string	Repository name
path	string	.ets file path
language	string	Programming language (arkts)
identifier	string	Function identifier / name
docstring	string	Cleaned, normalized function documentation
function	string	Original function source code
ast_function	string	AST representation generated with tree-sitter-arkts
obf_function	string	Legacy obfuscated-function field; currently equal to function
url	string	Repository or source-code URL when available
function_sha	string	Function-content identifier
source	string	Code source platform (GitHub / Gitee / GitCode)

Usage

from datasets import load_dataset

# Repository-isolated configuration
dataset = load_dataset(
    "hreyulog/arkts-code-docstring-pro",
    name="repo_split_balanced",
)

# Random-shuffle configuration
random_dataset = load_dataset(
    "hreyulog/arkts-code-docstring-pro",
    name="random_shuffle",
)

print(dataset["train"][0])
print(dataset["train"].features)

License Configuration

All included repositories were selected from locally identified open-source repositories. Users must still comply with the individual upstream repository licenses and attribution requirements. This Pro dataset does not grant a new license over the source code.

Citation

If you use this dataset in your research, please cite the following paper:

@misc{he2026arktscodesearchopensourcearktsdataset,
      title={ArkTS-CodeSearch: A Open-Source ArkTS Dataset for Code Retrieval},
      author={Yulong He and Artem Ermakov and Sergey Kovalchuk and Artem Aliev and Dmitry Shalymov},
      year={2026},
      eprint={2602.05550},
      archivePrefix={arXiv},
      primaryClass={cs.SE},
      url={https://arxiv.org/abs/2602.05550},
}