README.md

metadata

dataset_info:
  features:
    - name: nwo
      dtype: string
    - name: sha
      dtype: string
    - name: path
      dtype: string
    - name: language
      dtype: string
    - name: identifier
      dtype: string
    - name: docstring
      dtype: string
    - name: function
      dtype: string
    - name: ast_function
      dtype: string
    - name: obf_function
      dtype: string
    - name: url
      dtype: string
    - name: function_sha
      dtype: string
    - name: source
      dtype: string
  splits:
    - name: train
      num_bytes: 162123504
      num_examples: 19561
    - name: validation
      num_bytes: 18843858
      num_examples: 2445
    - name: test
      num_bytes: 19867797
      num_examples: 2446
  download_size: 42313088
  dataset_size: 200835159
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: validation
        path: data/validation-*
      - split: test
        path: data/test-*
task_categories:
  - text-retrieval
language:
  - en
  - zh

ArkTS-CodeSearch: A Open-Source ArkTS Dataset for Code Retrieval

Paper | Evaluation Code | Fine-Tuned Model

This dataset collects function-level information from ArkTS (HarmonyOS Ark TypeScript) projects, including original functions, docstrings, abstract syntax tree (AST) representations, obfuscated versions, and source code metadata. It is suitable for tasks such as code analysis, code understanding, AST research, and code search.

ArkTS is a core programming language in the OpenHarmony ecosystem. This dataset was constructed from open-source repositories crawled from GitHub and Gitee, targeting code retrieval and code evaluation tasks.

Note: ⚠️ Not Enterprise-Ready: This dataset is designed for research only. Some of the collected code may come from repositories without an explicit open-source license. Users should be aware that usage of such code may be subject to copyright restrictions.

Dataset Structure

The dataset contains three splits:

• train: Training set
• validation: Validation set
• test: Test set

Each split is a JSON Lines (.jsonl) file, where each line is a JSON object representing a single function.

Features / Columns

Field	Type	Description
nwo	string	Repository name
sha	string	Commit SHA
path	string	File path
language	string	Programming language
identifier	string	Function identifier / name
docstring	string	Function docstring
function	string	Original function source code
ast_function	string	AST representation of the function
obf_function	string	Obfuscated function source code
url	string	URL to the code in the repository
function_sha	string	Function-level SHA
source	string	Code source (GitHub / Gitee)

Usage

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("hreyulog/arkts-code-docstring")

# Inspect the first training example
print(dataset["train"][0])

# Check dataset features
print(dataset["train"].features)

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}, 
}