Datasets:

Modalities:
Text
Formats:
parquet
Libraries:
Datasets
Dask
cedricbonhomme's picture
Update README.md
16cedaf verified
metadata
dataset_info:
  features:
    - name: id
      dtype: string
    - name: title
      dtype: string
    - name: description
      dtype: string
    - name: patches
      list:
        - name: commit_message
          dtype: string
        - name: patch_text_b64
          dtype: string
        - name: url
          dtype: string
    - name: cwe
      dtype: string
  splits:
    - name: train
      num_bytes: 1619279339
      num_examples: 14510
    - name: test
      num_bytes: 180006724
      num_examples: 1613
  download_size: 825880045
  dataset_size: 1799286063
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: test
        path: data/test-*

Description

This dataset, CIRCL/vulnerability-cwe-patch, provides structured real-world vulnerabilities enriched with CWE identifiers and actual patches from platforms like GitHub and GitLab. It was built to support the development of tools for vulnerability classification, triage, and automated repair. Each entry includes metadata such as CVE/GHSA ID, a description, CWE categorization, and links to verified patch commits with associated diff content and commit messages.

The dataset is automatically extracted using a pipeline that fetches vulnerability records from several sources, filters out entries without patches, and verifies patch links for accessibility. Extracted patches are fetched, encoded in base64, and stored alongside commit messages for training and evaluation of ML models. Source Data

The dataset comprises 16,123 vulnerabilities and 18,810 associated patches. For training, we consider only those patches corresponding to vulnerabilities annotated with at least one CWE.

How to use with datasets

>>> import json
>>> from datasets import load_dataset

>>> dataset = load_dataset("CIRCL/vulnerability-cwe-patch")

>>> vulnerabilities = ["CVE-2025-60249", "CVE-2025-32413"]

>>> filtered_entries = dataset.filter(lambda elem: elem["id"] in vulnerabilities)

>>> for entry in filtered_entries["train"]:
...     print(entry["cwe"])
...     for patch in entry["patches"]:
...         print(f"  {patch['commit_message']}")
...
CWE-79 Improper Neutralization of Input During Web Page Generation (XSS or 'Cross-site Scripting')     
  [PATCH] fix: [security] Fixed a stored XSS vulnerability in user bios. Thanks to Dawid Czarnecki for reporting the issue.
CWE-79 Improper Neutralization of Input During Web Page Generation (XSS or 'Cross-site Scripting')
  [PATCH] fix: [security] sanitize user input in comments, bundles, and sightings - Escaped untrusted data in templates and tables to prevent XSS - Replaced unsafe innerHTML assignments with createElement/textContent - Encoded dynamic URLs using encodeURIComponent - Improved validation in Comment, Bundle, and Sighting models Credit: @Wachizungu

Schema

Each example contains:

- id: Vulnerability identifier (e.g., CVE-2023-XXXX, GHSA-XXXX)

- title: Human-readable title of the vulnerability

- description: Detailed vulnerability description

- patches: List of patch records, each with:

    url: Verified patch URL (GitHub/GitLab)

    patch_text_b64: Base64-encoded unified diff

    commit_message: Associated commit message

- cwe: List of CWE identifiers and names

The vulnerabilities can be sourced from:

- NVD CVE List — enriched with commit references

- GitHub Security Advisories (GHSA)

- GitLab advisories

- CSAF feeds from vendors including Red Hat, Cisco, and CISA

Use Cases

The dataset supports a range of security-focused machine learning tasks:

* Vulnerability classification

* CWE prediction from descriptions

* Patch generation from natural language

* Commit message understanding

Associated Code

The dataset is generated with the extraction pipeline from vulnerability-lookup/ML-Gateway, which includes logic for fetching, filtering, validating, and encoding patch data.