README.md

metadata

language:
  - en
tags:
  - vulnerability-detection
  - cve
  - code-changes
  - software-security
  - stratified-split
license: mit
dataset_info:
  features:
    - name: idx
      dtype: int64
    - name: func_before
      dtype: string
    - name: Vulnerability Classification
      dtype: string
    - name: vul
      dtype: int64
    - name: func_after
      dtype: string
    - name: patch
      dtype: string
    - name: CWE ID
      dtype: string
    - name: lines_before
      dtype: string
    - name: lines_after
      dtype: string
  splits:
    - name: train
      num_examples: 150909
    - name: validation
      num_examples: 18864
    - name: test
      num_examples: 18863
  dataset_original_file_size: 10GB uuncompressed

MSR Data Cleaned - C/C++ Code Vulnerability Dataset

📌 Dataset Description

A curated collection of C/C++ code vulnerabilities paired with:

• CVE details (scores, classifications, exploit status)
• Code changes (commit messages, added/deleted lines)
• File-level and function-level diffs

🔍 Sample Data Structure from original file

+---------------+-----------------+----------------------+---------------------------+
| CVE ID        | Attack Origin   | Publish Date         | Summary                   |
+===============+=================+======================+===========================+
| CVE-2015-8467 | Remote          | 2015-12-29           | "The samldb_check_user..."|
+---------------+-----------------+----------------------+---------------------------+
| CVE-2016-1234 | Local           | 2016-01-15           | "Buffer overflow in..."   |
+---------------+-----------------+----------------------+---------------------------+

Note: This is a simplified preview; the full dataset includes additional fields like commit_id, func_before, etc.

1. Accessing in Colab

!pip install huggingface_hub -q
from huggingface_hub import snapshot_download

repo_id = "starsofchance/MSR_data_cleaned"
dataset_path = snapshot_download(repo_id=repo_id, repo_type="dataset")

2. Extracting the Dataset

!apt-get install unzip -qq
!unzip "/root/.cache/huggingface/.../MSR_data_cleaned.zip" -d "/content/extracted_data"

**Note: Extracted size is 10GB (1.5GB compressed). Ensure sufficient disk space.

3. Creating Splits (Colab Pro Recommended)

We used this memory-efficient approach:

from datasets import load_dataset
dataset = load_dataset("csv", data_files="MSR_data_cleaned.csv", streaming=True)

# Randomly distribute rows (80-10-10)
for row in dataset:
    rand = random.random()
    if rand < 0.8: write_to(train.csv)
    elif rand < 0.9: write_to(validation.csv)
    else: write_to(test.csv)

Hardware Requirements:

• Minimum 25GB RAM
• Strong CPU (Colab Pro T4 GPU recommended)

##📊 Dataset Statistics

• Number of Rows: 188,636
• Vulnerability Distribution:
  • Vulnerable (1): 18,863 (~10%)
  • Non-Vulnerable (0): 169,773 (~90%)

##📋 Data Fields Description

• CVE_ID: Unique identifier for the vulnerability (Common Vulnerabilities and Exposures).
• CWE_ID: Weakness category identifier (Common Weakness Enumeration).
• Score: CVSS score indicating severity (float, 0-10).
• Summary: Brief description of the vulnerability.
• commit_id: Git commit hash linked to the code change.
• codeLink: URL to the code repository or commit.
• file_name: Name of the file containing the vulnerability.
• func_after: Function code after the change.
• lines_after: Code lines after the change.
• Access_Gained: Type of access gained by exploiting the vulnerability.
• Attack_Origin: Source of the attack (e.g., Remote, Local).
• Authentication_Required: Whether authentication is needed to exploit.
• Availability: Impact on system availability.
• CVE_Page: URL to the CVE details page.
• Complexity: Complexity of exploiting the vulnerability.
• Confidentiality: Impact on data confidentiality.
• Integrity: Impact on data integrity.
• Known_Exploits: Details of known exploits, if any.
• Publish_Date: Date the vulnerability was published.
• Update_Date: Date of the last update to the vulnerability data.
• Vulnerability_Classification: Type or category of the vulnerability.
• add_lines: Lines added in the commit.
• del_lines: Lines deleted in the commit.
• commit_message: Description of the commit.
• files_changed: List of files modified in the commit.
• func_before: Function code before the change.
• lang: Programming language (e.g., C, C++).
• lines_before: Code lines before the change.

splits file for UltiVul project:

🔍 Sample Data Structure (from train.csv)

{
 'idx': 0, # Unique ID within the train split
 'func_before': '...', # String containing function code before change
 'Vulnerability Classification': '...', # Original vulnerability type classification
 'vul': 0, # Integer: 0 for non-vulnerable, 1 for vulnerable (target label)
 'func_after': '...', # String containing function code after change
 'patch': '...', # String containing diff patch
 'CWE ID': '...', # String CWE ID, e.g., "CWE-119"
 'lines_before': '...', # String lines before change context
 'lines_after': '...' # String lines after change context
}

**Note: This shows the structure of the final split files (train.csv, validation.csv, test.csv). The original MSR_data_cleaned.csv contains many more metadata fields.

##📦 Dataset New Files The dataset is available as three CSV files (specially created for the UltiVul project) hosted on Hugging Face, uploaded via huggingface_hub:

• train.csv Size: 667 MB Description: Training split with 150,909 samples, approximately 80% of the data.
• validation.csv Size: 86 MB Description: Validation split with 18,864 samples, approximately 10% of the data.
• test.csv Size: 84.8 MB Description: Test split with 18,863 samples, approximately 10% of the data.

🙏 Acknowledgements Original dataset provided by Fan et al., 2020 Thanks to the Hugging Face team for dataset hosting tools.

📜 Citation

@inproceedings{fan2020ccode,
  title={A C/C++ Code Vulnerability Dataset with Code Changes and CVE Summaries},
  author={Fan, Jiahao and Li, Yi and Wang, Shaohua and Nguyen, Tien N},
  booktitle={MSR '20: 17th International Conference on Mining Software Repositories},
  pages={1--5},
  year={2020},
  doi={10.1145/3379597.3387501}
}

🌟 Dataset Creation

• Source: Original data from MSR 2020 Paper
• Processing:
  • Cleaned and standardized CSV format
  • Stream-based splitting to handle large size
  • Preserved all original metadata