RISys-Lab/Benchmarks_CyberSec_SECURE

Dataset Card for SECURE (RISys-Lab Mirror)

⚠️ Disclaimer: > This repository is a mirror/re-host of the original SECURE benchmark.
RISys-Lab is not the author of this dataset. We are hosting this copy in Parquet format to ensure seamless integration and stability for our internal evaluation pipelines. All credit belongs to the original authors listed below.

Repository Intent

This Hugging Face dataset is a re-host of the original SECURE benchmark. It has been converted to Parquet format to support efficient loading and configuration handling in the datasets library. If you are looking for the official release, please visit the Original GitHub Repository.

Dataset Summary

SECURE (Security Extraction, Understanding & Reasoning Evaluation) is a benchmark designed to assess Large Language Models (LLMs) in realistic cybersecurity scenarios, with a specific focus on Industrial Control Systems (ICS).

The benchmark consists of six distinct datasets that evaluate knowledge extraction, understanding, and reasoning based on industry-standard sources (such as CISA advisories and MITRE frameworks).

Supported Tasks

The dataset is divided into six configurations, each targeting a specific skill:

1. MAET (MITRE ATT&CK Extraction Task): Multiple-choice questions mapping attack behaviors to MITRE ATT&CK techniques.
2. CWET (Common Weakness Extraction Task): Multiple-choice questions identifying Common Weaknesses (CWEs) from descriptions.
3. KCV (Knowledge test on Common Vulnerabilities): A knowledge verification task for CVEs.
4. VOOD (Vulnerability Out-of-Distribution): A task designed to test model performance on out-of-distribution vulnerability data.
5. RERT (Risk Evaluation Reasoning Task): Evaluating the model's ability to reason about risk based on vulnerability overviews (e.g., from CISA ICS advisories).
6. CPST (CVSS Problem Solving Task): A regression/reasoning task where the model must determine the CVSS (Common Vulnerability Scoring System) score.

Dataset Structure

Data Splits & Configurations

The dataset is organized into 6 configurations.

Important Note on Validation Splits: > The original SECURE benchmark provided test sets. To facilitate few-shot evaluation in our pipeline, we randomly sampled 5 examples from the original data to create a val split for each configuration.

Config Name	Full Task Name	Validation Size (Few-Shot)	Test Size
MAET	MITRE ATT&CK Extraction Task	5	1,067
CWET	Common Weakness Extraction Task	5	959
KCV	Knowledge on Common Vulnerabilities	5	461
VOOD	Vulnerability Out-of-Distribution	5	461
RERT	Risk Evaluation Reasoning Task	5	995
CPST	CVSS Problem Solving Task	5	95

Data Fields

MAET, CWET (Multiple Choice)

• url (string): Source URL.
• prompt (string): The full input prompt.
• question (string): The specific question text.
• options (sequence): A list/sequence of answer choices (e.g., ["Option A...", "Option B..."]).
• answer (string): The correct option (e.g., "A").

KCV, VOOD (True / False)

• url (string): Source URL.
• prompt (string): The full input prompt.
• question (string): The question text.
• answer (string): The correct answer string.

RERT (Reasoning)

• source-url (string): Source URL (e.g., CISA advisory).
• prompt (string): The input prompt requesting a risk evaluation.
• vulnerability-overview (string): Context describing the vulnerability.
• answer (string): The gold-standard risk evaluation text.

CPST (Scoring)

• prompt (string): The input prompt containing vulnerability details.
• cvss-v3-vector-string (string): The CVSS vector string (e.g., CVSS:3.1/AV:N/AC:L...).
• answer (float64): The correct CVSS Base Score (e.g., 7.5).

Usage

You can load a specific task using the configuration name.

from datasets import load_dataset

# Load the MITRE ATT&CK (MAET) test set
dataset = load_dataset("RISys-Lab/Benchmarks_CyberSec_SECURE", "MAET", split="test")

# Load the few-shot examples (val split)
few_shot_examples = load_dataset("RISys-Lab/Benchmarks_CyberSec_SECURE", "MAET", split="val")

# Access an example
print(dataset[0])

Additional Information

Original Authors

The dataset was developed by:

• Dipkamal Bhusal
• Nidhi Rastogi
• Md Tanvirul Alam
• (and contributors from Rochester Institute of Technology)

Citation

Please cite the original ArXiv paper if you use this dataset:

@misc{bhusal2024securebenchmarkinglargelanguage,
      title={SECURE: Benchmarking Large Language Models for Cybersecurity Advisory}, 
      author={Dipkamal Bhusal and Nidhi Rastogi and Md Tanvirul Alam and Le Nguyen and Xashru Shrestha and Qiben Yan and Rui Li and Tuan Vu and Nathan Lewis and Y. S. Rao},
      year={2024},
      eprint={2405.20441},
      archivePrefix={arXiv},
      primaryClass={cs.CR},
      url={https://arxiv.org/abs/2405.20441}, 
}

License

An explicit license file was not found in the original repository. This mirror is provided for research purposes. All rights remain with the original authors.