Younis2003/secure-code-repair-sft

Secure Code Repair (SFT)

📌 Overview

Secure Code Repair (SFT) is a high-quality supervised fine-tuning dataset designed to train large language models to repair vulnerable source code by transforming insecure implementations into secure and corrected versions.

Each sample consists of real-world vulnerable code paired with its corresponding secure fix. The dataset has been carefully engineered using a strict validation and filtering pipeline to ensure security relevance, correctness, and training stability.

This dataset represents the first training stage (Supervised Fine-Tuning) for secure code generation systems.

---

🎯 Motivation

Large Language Models often generate code that is syntactically correct but security-wise unsafe.
The goal of this dataset is to:

• Teach models how to fix vulnerabilities, not just detect them
• Encourage secure coding patterns across multiple languages
• Provide clean and validated examples suitable for academic research and applied security systems

---

🗂️ Data Sources

This dataset is derived from:

• CVEFixes — a curated collection of real-world vulnerability patches linked to CVEs

All samples were processed using a custom data engineering pipeline that removes:

• Incomplete or empty code examples
• Trivial or cosmetic-only fixes
• Destructive fixes that remove functionality instead of securing it
• Low-quality or inconsistent language samples

---

🧩 Supported Programming Languages

This dataset supports vulnerable-to-secure code transformations for the following programming languages:

• C
• C++
• Python
• Java
• JavaScript
• PHP
• Go
• Ruby
• C#

Only allow-listed languages are included during preprocessing to ensure consistency, reliability, and high-quality training data.

---

🧱 Dataset Structure

Each sample follows the schema below:

{
  "id": "CVE-XXXX",
  "language": "C",
  "input": {
    "vulnerable_code": "..."
  },
  "output": {
    "secure_code": "..."
  },
  "metadata": {
    "source": "CVEFixes"
  }
}