Code2Doc: Function-Documentation Pairs Dataset

A curated dataset of 13,358 high-quality function-documentation pairs extracted from popular open-source repositories on GitHub. Designed for training models to generate documentation from code.

Dataset Description

This dataset contains functions paired with their docstrings/documentation comments from 5 programming languages, extracted from well-maintained, highly-starred GitHub repositories.

Languages Distribution

Language	Train	Val	Test	Total
Java	6,560 (61.4%)	820	820	8,200
Python	2,885 (27.0%)	360	362	3,607
TypeScript	681 (6.4%)	85	86	852
JavaScript	428 (4.0%)	53	55	536
C++	130 (1.2%)	16	17	163
Total	10,684	1,334	1,340	13,358

Source Repositories

The data was extracted from high-quality open-source projects including:

Python: Django, PyTorch, Pandas, NumPy, scikit-learn, FastAPI, Flask, Celery, Airflow, Requests

Java: Guava, Elasticsearch, Spring Framework, Spring Boot, Apache Kafka, Commons-Lang

TypeScript: TypeScript, VS Code, Angular, Prisma, Grafana, Storybook, NestJS

JavaScript: React, Node.js, Lodash, Axios, Express

C++: OpenCV, Protobuf, Folly, gRPC, LLVM, TensorFlow

Dataset Structure

Data Fields

Field	Type	Description
function_name	string	Name of the function/method
function_code	string	Complete source code of the function
documentation	string	Extracted docstring/documentation
language	string	Programming language
file_path	string	Original file path in repository
line_number	int	Line number where function starts
parameters	list[string]	List of parameter names
return_type	string	Return type annotation (if available)
has_type_hints	bool	Whether function has type annotations
complexity	int	Cyclomatic complexity score
quality_score	float	Documentation quality score (0-10)
repo_name	string	Source repository (owner/repo)
repo_stars	int	Repository star count at extraction time
docstring_style	string	Documentation style (google, numpy, sphinx, jsdoc, javadoc, doxygen)
is_async	bool	Whether function is async

Data Splits

• Train: 10,684 samples (80%)
• Validation: 1,334 samples (10%)
• Test: 1,340 samples (10%)

Splits are stratified by language to maintain consistent distribution across sets.

Data Processing Pipeline

The dataset was created through a multi-stage pipeline:

1. Extraction: Used tree-sitter parsers to accurately extract functions with documentation
2. Basic Filtering: Removed test functions, trivial functions, and applied length constraints
3. Quality Scoring: Scored documentation completeness (parameters, returns, examples)
4. Deduplication: Removed exact and near-duplicates using MinHash LSH
5. AI Detection: Filtered potentially AI-generated documentation

Quality Criteria

• Minimum documentation length: 20 characters
• Maximum documentation length: 10,000 characters
• Minimum code length: 50 characters
• Excluded test functions and trivial getters/setters
• Required meaningful documentation structure

Usage

from datasets import load_dataset

dataset = load_dataset("kaanrkaraman/code2doc")

# Access splits
train_data = dataset["train"]
val_data = dataset["val"]
test_data = dataset["test"]

# Example: Get a Python function
python_samples = train_data.filter(lambda x: x["language"] == "python")
sample = python_samples[0]

print(f"Function: {sample['function_name']}")
print(f"Code:\n{sample['function_code']}")
print(f"Documentation:\n{sample['documentation']}")

For Fine-tuning

def format_for_training(example):
    return {
        "input": f"Generate documentation for the following {example['language']} function:\n\n{example['function_code']}",
        "output": example["documentation"]
    }

formatted_dataset = dataset.map(format_for_training)

Intended Use

• Training code documentation generation models
• Fine-tuning LLMs for code-to-text tasks
• Evaluating documentation quality metrics
• Research on code understanding and generation

Limitations

• Heavily weighted towards Java due to verbose documentation practices
• C++ representation is small due to different documentation conventions
• Documentation quality varies by repository coding standards
• Extracted from a specific snapshot in time (December 2025)

Citation

@misc{recep_kaan_karaman_2025,
  author       = {Recep Kaan Karaman and Meftun Akarsu},
  title        = {code2doc (Revision cadd4e4)},
  year         = 2025,
  url          = {https://huggingface.co/datasets/kaanrkaraman/code2doc},
  doi          = {10.57967/hf/7310},
  publisher    = {Hugging Face}
}

License

This dataset is released under the CC BY 4.0 License. The source code comes from repositories with permissive licenses (MIT, Apache 2.0, BSD).