prabhatl0dhi/C_C_plus_plus_Patch_Dataset_for_Secure_Code_LLMs

C/C++ Security Vulnerability & Automated Patch Dataset for llms fine tuning

A large-scale corpus of 23,826 function-level vulnerability repair pairs ($Vulnerable \rightarrow Fixed$) mined from 64 major C/C++ open-source projects spanning 12,130 unique commits identified by the security-focused mining pipeline.

This dataset is specifically structured for Automated Program Repair (APR), Vulnerability Assessment, and Instruction Fine-Tuning of Code LLMs (e.g., Qwen2.5-Coder, DeepSeek-Coder, CodeLlama).

---

Dataset Summary

• Total Samples: 23,826 function pairs
• Source Repositories: 64 curated open-source C/C++ codebases
• Unique Commits: 12,130 identified by the security-focused mining pipeline
• Median Patch Size: +3 lines added, -1 line deleted
• Exact Duplicate Pairs: 0 according to pair_hash (23,790 unique vulnerable functions, 23,810 unique fixed functions)
• Token Distribution (BPE - Qwen2.5-Coder):
  • Vulnerable Code: Median 363 tokens (P95: 1,993)
  • Fixed Code: Median 379 tokens (P95: 2,026)
  • Full SFT Training Prompt: Median 846 tokens (P95: 4,116; 94.95% fit in $\le 4,096$)

---

Mining Methodology

The dataset was mined using an automated git-mining and static AST function extraction pipeline:

1. Repository Selection: 64 active, mission-critical C/C++ codebases across systems programming, cryptography, operating systems, embedded firmware, networking, virtualization, and parsers.
2. Commit Filtering: Commits were identified using security-focused regex patterns in commit messages (e.g., CVE IDs, memory corruption taxonomy terms, security advisories).
3. Function Extraction: For each modified .c, .cpp, .cc, or .h file in a security commit, the pipeline used AST/heuristic function parsing to extract the exact enclosing function before (vulnerable_code) and after (fixed_code) the commit.
4. Structural Validation: Multi-file diffs and refactorings were isolated to function boundaries. Function pairs where syntax was unextractable or unchanged were discarded.

---

Dataset Schema & Field Descriptions

Each row in dataset_vuln_pairs.jsonl is a JSON object with the following fields:

Field Name	Type	Description
pair_hash	string	Unique SHA-256 hash computed over normalized code: hashlib.sha256(f"{vulnerable_code.strip()}|||{fixed_code.strip()}".encode("utf-8")).hexdigest().
repo_name	string	GitHub repository identifier (e.g., openssl/openssl, torvalds/linux, redis/redis).
commit_hash	string	Full Git commit SHA where the patch was applied.
commit_message	string	Complete commit message from the developer explaining the patch.
commit_date	string	ISO 8601 timestamp of the commit.
file_path	string	Relative path to the modified file within the repository.
function_name	string	Name of the extracted C/C++ function.
vulnerable_code	string	Complete source code of the function before the security fix.
fixed_code	string	Complete source code of the function after the security fix.
vuln_category	string	Primary vulnerability taxonomy classified from commit semantics (e.g., Buffer Overflow, Memory Leak, Use After Free, Null Pointer Dereference).
cve_ids	list[string]	List of explicit Common Vulnerabilities and Exposures IDs mentioned in the commit (e.g., ["CVE-2023-0286"]).
source_url	string	Direct permalink to the commit on GitHub.

---

Known Limitations & Usage Recommendations

1. Mined Signals vs. Ground Truth: Vulnerability categories and security designations are derived from developer commit messages and metadata (weak supervision). While indicative of security-relevant code modifications, they have not undergone manual line-by-line verification by security researchers.
2. Backport & Sibling Repository Clusters: Sibling projects and forks (e.g., openssl $\leftrightarrow$ boringssl $\leftrightarrow$ libressl, or tcpdump $\leftrightarrow$ libpcap) share commit histories and backported patches. Users should perform group-aware / cluster-aware splits to ensure all sibling repositories reside together in either the train or test partition.
3. Commit Co-Occurrence: 32.3% of commits modified more than one function. When creating train/val/test splits, always split by Commit SHA or Repository, never by individual row indices.
4. Token Truncation: For models with a 4,096 token context window, 94.95% of pairs fit without truncation. For ~99.85% coverage, use an 8,192-token context window (only 36 outliers exceed 8k).

---

Quickstart Usage

import json

def load_dataset(filepath="dataset_vuln_pairs.jsonl"):
    with open(filepath, "r", encoding="utf-8") as f:
        for line in f:
            yield json.loads(line)

# Example: Inspect Gold-Tier CVE samples
for sample in load_dataset():
    if sample["cve_ids"]:
        print(f"[{sample['repo_name']}] {sample['cve_ids']} - {sample['function_name']}")
        print(f"Commit: {sample['commit_message'].splitlines()[0]}")
        break

---

Provenance & License

• Dataset Aggregation & Tooling License: MIT License. The MIT license covers the dataset packaging, schemas, mining tooling, and documentation.
• Source Code Snippet Provenance: The underlying C/C++ function snippets originate from public open-source projects hosted on GitHub and remain subject to their respective upstream licenses (e.g., GPL-2.0, Apache-2.0, MIT, BSD, MPL). The dataset aggregation license does not alter or supersede the upstream license terms of individual code snippets. Users must adhere to upstream licenses when using these code snippets.