README.md

metadata

license: mit
dataset_info:
  features:
    - name: idx
      dtype: int64
    - name: project
      dtype: string
    - name: project_url
      dtype: string
    - name: filepath
      dtype: string
    - name: commit_id
      dtype: string
    - name: commit_message
      dtype: string
    - name: is_vulnerable
      dtype: bool
    - name: hash
      dtype: string
    - name: func_name
      dtype: string
    - name: func_body
      dtype: string
    - name: changed_lines
      dtype: string
    - name: changed_statements
      dtype: string
    - name: cve_list
      sequence: string
    - name: cwe_list
      sequence: string
    - name: fixed_func_idx
      dtype: int64
    - name: context
      struct:
        - name: Execution Environment
          sequence: string
        - name: Explanation
          sequence: string
        - name: External Function
          sequence: string
        - name: Function Argument
          sequence: string
        - name: Globals
          sequence: string
        - name: Type Execution Declaration
          sequence: string
  splits:
    - name: train
      num_bytes: 119514243
      num_examples: 25440
  download_size: 30875803
  dataset_size: 119514243
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

Dataset Card for Dataset Name

SecVulEval is a collection of real-world C/C++ vulnerabilities.

Dataset Details

Dataset Description

The dataset is curated by collecting C/C++ vulnerability from NVD. It features statement-level vulnerable information, context information for vulnerable functions (is_vulnerable=True), and other metadata such as CVE, CWE, commit information. The dataset contains vulnerable and non-vulnerable function samples.

Dataset Sources

The vulnerabilities (CVEs) are collected from NVD (https://nvd.nist.gov). Then, the corresponding patches to the vulnerabilities are collected from their respective git repositories.

Uses

The dataset comprises both vulnerable (43.23%) and non-vulnerable (56.77%) functions, with a total collection of 25,440 function. This large collection of functions make it suitable for training vulnerability detection model. The statement-level info, along with contextual information can make context-aware detection at finer-grained level possible. The dataset can also be used to evaluate C/C++ vulnerability detection models.

Dataset Structure

The dataset has 15 different fields.

• The project_url column has 735 different values while the project column has 707 unique values. This is because for project == "Android", there are multiple different repositories.
• The changed_lines and changed_statements columns include the changes in made in the patch as a list of (line, code) pair. Vulnerable functions include the deleted lines/statements and the non-vulnerable functions has the added lines/statements.
• Some functions/vulnerabilities can be assigned to more than one CVE/CWE which is why cve_list and cwe_list are given as lists, although in most cases there would be only one CVE and CWE id.
• The fixed_func_id includes the idx number (first field in the dataset) of the corresponding fixed patch of a vulnerable function. This helps to easily pair the vulnerable functions with their fixing code. For non-vulnerable code it doesn't make sense for a "fixed" version and the fixed_func_id is just itself.
• The context field includes contextual information for vulnerable functions according to the five categories as discussed in the paper. It is added as the list of symbols and an explanation as generated by the LLM.

Other fields are self-explanatory.