MultiLang Code Parser Dataset (MLCPD)

MultiLang-Code-Parser-Dataset (MLCPD) provides a large-scale, unified dataset of parsed source code across 10 major programming languages, represented under a universal schema that captures syntax, semantics, and structure in a consistent format.

Each entry corresponds to one parsed source file and includes:

• Language metadata
• Code-level statistics (lines, errors, AST nodes)
• Universal Schema JSON (normalized structural representation)

MLCPD enables robust cross-language analysis, code understanding, and representation learning by providing a consistent, language-agnostic data structure suitable for both traditional ML and modern LLM-based workflows.

---

📂 Dataset Structure

MultiLang-Code-Parser-Dataset/
├── c_parsed_1.parquet
├── c_parsed_2.parquet
├── c_parsed_3.parquet
├── c_parsed_4.parquet
├── c_sharp_parsed_1.parquet
├── ...
└── typescript_parsed_4.parquet

Each file corresponds to one partition of a language (~175k rows each).

Each record contains:

---

📊 Key Statistics

The dataset is clean, lossless, and statistically balanced across languages.
It offers both per-language and combined cross-language representations.

---

🚀 Use Cases

MLCPD can be directly used for:

• Cross-language code representation learning
• Program understanding and code similarity tasks
• Syntax-aware pretraining for LLMs
• Code summarization, clone detection, and bug prediction
• Graph-based learning on universal ASTs
• Benchmark creation for cross-language code reasoning

---

🔍 Features

• Universal Schema: A unified structural representation harmonizing AST node types across languages.
• Compact Format: Stored in Apache Parquet, allowing fast access and efficient querying.
• Cross-Language Compatibility: Enables comparative code structure analysis across multiple programming ecosystems.
• Error-Free Parsing: 99.9999% successful schema conversions across ~7M code files.
• Statistical Richness: Includes per-language metrics such as mean line count, AST size, and error ratios.
• Ready for ML Pipelines: Compatible with PyTorch, TensorFlow, Hugging Face Transformers, and graph-based models.

---

📥 How to Access the Dataset

Using the Hugging Face datasets Library

This dataset is hosted on the Hugging Face Hub and can be easily accessed using the datasets library.

Install the Required Library

pip install datasets

Import Library

from datasets import load_dataset

Load the Entire Dataset

dataset = load_dataset(
    "jugalgajjar/MultiLang-Code-Parser-Dataset"
)

Load a Specific Language File

dataset = load_dataset(
    "jugalgajjar/MultiLang-Code-Parser-Dataset",
    data_files="python_parsed_1.parquet"
)

Stream Data

dataset = load_dataset(
    "jugalgajjar/MultiLang-Code-Parser-Dataset",
    data_files="python_parsed_1.parquet",
    streaming=True
)

Access Data Content (After Downloading)

try:
    for example in dataset["train"].take(5):
        print(example)
        print("-"*25)
except Exception as e:
    print(f"An error occurred: {e}")

Manual Download

You can also manually download specific language files from the Hugging Face repository page:

1. Visit https://huggingface.co/datasets/jugalgajjar/MultiLang-Code-Parser-Dataset
2. Navigate to the Files tab
3. Click on the language file you want (e.g., python_parsed_1.parquet)
4. Use the Download button to save locally

---

🧾 Citation

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

Gajjar, J., & Subramaniakuppusamy, K. (2025).
MLCPD: A Unified Multi-Language Code Parsing Dataset with Universal AST Schema.
arXiv preprint arXiv:2510.16357

@article{gajjar2025mlcpd,
  title={MLCPD: A Unified Multi-Language Code Parsing Dataset with Universal AST Schema},
  author={Gajjar, Jugal and Subramaniakuppusamy, Kamalasankari},
  journal={arXiv preprint arXiv:2510.16357},
  year={2025}
}

---

📜 License

This dataset is released under the MIT License.
You are free to use, modify, and redistribute it for research and educational purposes, with proper attribution.

---

🙏 Acknowledgements

• StarCoder Dataset for source code samples
• TreeSitter for parsing
• Hugging Face for dataset hosting

---

📧 Contact

For questions, collaborations, or feedback:

• Primary Author: Jugal Gajjar
• Email: 812jugalgajjar@gmail.com
• LinkedIn: linkedin.com/in/jugal-gajjar/

---

⭐ If you find this dataset useful, consider liking the dataset and the GitHub repository and sharing your work that uses it.