README.md

metadata

annotations_creators:
  - machine-generated
language_creators:
  - found
language:
  - code
  - zh
  - en
license: other
multilinguality:
  - multilingual
pretty_name: JihuLab Code Dataset
size_categories:
  - 1M<n<10M
source_datasets:
  - original
task_categories:
  - text-generation
tags:
  - code
  - chinese
configs:
  - config_name: default
    data_files:
      - split: train
        path: data.parquet
    default: true
dataset_info:
  features:
    - name: code
      dtype: string
    - name: repo_name
      dtype: string
    - name: path
      dtype: string
    - name: language
      dtype: string
    - name: license
      dtype: string
    - name: size
      dtype: int64

JihuLab Code Dataset

Dataset Description

This dataset was compiled from code repositories hosted on JihuLab, a GitLab-based code hosting platform operated by JiHu (GitLab's Chinese joint venture). JihuLab serves as the primary GitLab instance for Chinese developers and enterprises, offering localized services and compliance with Chinese regulations. This dataset is particularly valuable for training code models with Chinese language understanding and enterprise-level coding practices.

Dataset Summary

Statistic	Value
Total Files	1,853,253
Total Repositories	11,589
Total Size	1.5 GB (compressed Parquet) / 12.76 GB (uncompressed)
Programming Languages	304
File Format	Parquet with Zstd compression

Key Features

• Chinese developer ecosystem: Contains code from JihuLab, GitLab's official Chinese distribution, featuring Chinese comments, documentation, and variable names
• Diverse language coverage: Spans 304 programming languages identified by go-enry (based on GitHub Linguist rules)
• Rich metadata: Includes repository name, file path, detected language, license information, and file size
• Enterprise and open-source projects: Includes code from both individual developers and Chinese enterprises using GitLab
• Quality filtered: Extensive filtering to remove vendor code, build artifacts, generated files, and low-quality content

Languages

The dataset includes 304 programming languages. The top 30 languages by file count:

Rank	Language	File Count
1	Java	348,517
2	C	209,924
3	JavaScript	191,164
4	Python	172,798
5	C++	136,046
6	Go	80,000
7	TypeScript	79,067
8	HTML	69,173
9	C#	64,511
10	Rust	50,515
11	Shell	43,352
12	Vue	40,687
13	TSX	36,844
14	CSS	34,779
15	Makefile	26,227
16	Ruby	25,812
17	PHP	21,401
18	CMake	15,292
19	Kotlin	14,220
20	BitBake	13,060
21	SCSS	10,957
22	Scala	9,333
23	Dart	9,125
24	Lua	7,413
25	ASP.NET	7,005
26	Vim Script	5,710
27	Unix Assembly	5,239
28	Starlark	5,134
29	Objective-C	4,931
30	Factor	4,920

Licenses

The dataset includes files from repositories with various licenses. Repositories with restrictive licenses (CC-BY-ND variants, Commons Clause, SSPL) were excluded:

License	File Count
apache-2.0	551,008
unknown	535,320
mit	320,834
agpl-3.0	169,922
gpl-2.0	112,829
bsd	65,104
cc0-1.0	13,557
lgpl-3.0	12,871
lgpl-2.1	9,960
bsd-3-clause	9,109
bsl-1.1	8,972
epl-1.0	7,494
gpl-3.0	7,476
unlicense	6,265
cc-by-3.0	4,717
cc-by-nc-sa	4,339
mpl-2.0	3,847
cc-by-4.0	2,459
cc-by-nc-sa-4.0	1,715
cc-by-sa-4.0	1,701
bsd-2-clause	1,599
cc-by-nc-nd-4.0	1,222
isc	520
wtfpl	274
cc-by-nc-4.0	122
cc-by-sa	13
cc-by-sa-3.0	4

Dataset Structure

Data Fields

Field	Type	Description
code	string	Content of the source file (UTF-8 encoded)
repo_name	string	Name of the JihuLab repository (format: username/repo or group/subgroup/repo)
path	string	Path of the file within the repository (relative to repo root)
language	string	Programming language as identified by go-enry
license	string	License of the repository (SPDX identifier or "unknown")
size	int64	Size of the source file in bytes

Data Format

• Format: Apache Parquet with Zstd compression
• File Structure: Single consolidated file (data.parquet)

Data Splits

All examples are in the train split. There is no validation or test split.

Example Data Point

{
    'code': 'package com.example.demo;\n\nimport org.springframework.boot.*;\nimport org.springframework.boot.autoconfigure.*;\n...',
    'repo_name': 'SmallQ/demo',
    'path': 'src/main/java/com/example/demo/DemoApplication.java',
    'language': 'Java',
    'license': 'unknown',
    'size': 400
}

Dataset Creation

Pipeline Overview

The dataset was created through a multi-stage pipeline:

1. Repository Discovery: Paginated API requests to JihuLab's GitLab API (/api/v4/projects) to enumerate public repositories
2. Branch Selection: Using the repository's default branch (typically main or master)
3. Repository Downloading: Downloading repository archives via JihuLab's archive endpoint
4. Content Extraction: Extracting and filtering source code files
5. Parquet Generation: Writing filtered records to Parquet with Zstd compression

Language Detection

Programming languages are detected using go-enry, a Go port of GitHub's Linguist library. Only files classified as Programming or Markup language types are included (Data and Prose types are excluded).

License Detection

Licenses are detected by:

1. Scanning for license files (LICENSE, LICENSE.txt, LICENSE.md, COPYING, etc.)
2. Matching license text against known patterns (MIT, Apache 2.0, GPL variants, BSD, Creative Commons, etc.)
3. Defaulting to "unknown" if no license can be detected

Blocked Licenses: The following restrictive licenses are excluded from the dataset:

• cc-by-nd, cc-by-nd-2.0, cc-by-nd-3.0, cc-by-nd-4.0 (Creative Commons No-Derivatives)
• commons-clause
• sspl, sspl-1.0 (Server Side Public License)

File Filtering

Extensive filtering is applied to ensure data quality:

Size Limits

Limit	Value
Max repository ZIP size	48 MB
Max single file size	1 MB
Max line length	1,000 characters

Excluded Directories

• Configuration: .git/, .github/, .gitlab/, .vscode/, .idea/, .vs/, .settings/, .eclipse/, .project/, .metadata/
• Vendor/Dependencies: node_modules/, bower_components/, jspm_packages/, vendor/, third_party/, 3rdparty/, external/, packages/, deps/, lib/vendor/, target/dependency/, Pods/
• Build Output: build/, dist/, out/, bin/, target/, release/, debug/, .next/, .nuxt/, _site/, _build/, __pycache__/, .pytest_cache/, cmake-build-*, .gradle/, .maven/

Excluded Files

• Lock Files: package-lock.json, yarn.lock, pnpm-lock.yaml, Gemfile.lock, Cargo.lock, poetry.lock, Pipfile.lock, composer.lock, go.sum, mix.lock
• Minified Files: Any file containing .min. in the name
• Binary Files: .exe, .dll, .so, .dylib, .a, .lib, .o, .obj, .jar, .war, .ear, .class, .pyc, .pyo, .wasm, .bin, .dat, .pdf, .doc, .docx, .xls, .xlsx, .ppt, .pptx, .zip, .tar, .gz, .bz2, .7z, .rar, .jpg, .jpeg, .png, .gif, .bmp, .ico, .svg, .mp3, .mp4, .avi, .mov, .wav, .flac, .ttf, .otf, .woff, .woff2, .eot
• System Files: .DS_Store, thumbs.db

Content Filtering

• UTF-8 Validation: Files must be valid UTF-8 encoded text
• Binary Detection: Files detected as binary by go-enry are excluded
• Generated Files: Files with generation markers in the first 500 bytes are excluded:
  • generated by, do not edit, auto-generated, autogenerated, automatically generated, code generator, generated code, this file is generated, @generated, <auto-generated
• Empty Files: Files that are empty or contain only whitespace are excluded
• Long Lines: Files with any line exceeding 1,000 characters are excluded
• go-enry Filters: Additional filtering using go-enry's IsVendor(), IsImage(), IsDotFile(), IsTest(), and IsGenerated() functions
• Documentation-only Repos: Repositories containing only documentation files (no actual code) are skipped

Source Data

All data originates from public repositories hosted on JihuLab.

Considerations for Using the Data

Personal and Sensitive Information

The dataset may contain:

• Email addresses in code comments or configuration files
• API keys or credentials that were accidentally committed
• Personal information in comments or documentation

Users should exercise caution and implement appropriate filtering when using this data.

Licensing Information

This dataset is a collection of source code from repositories with various licenses. Any use of all or part of the code gathered in this dataset must abide by the terms of the original licenses, including attribution clauses when relevant. The license field in each data point indicates the license of the source repository.