AG-31625874/SemGuard-Dataset

SemGuard: Arabic Security Dataset (v2)

This dataset is released as part of the research paper: "SemGuard: A Triple-Anchor Semantic Security Gateway for Multilingual Prompt Attack Detection in Large Language Models".

Dataset Overview

The dataset provides a comprehensive benchmark for evaluating LLM prompt attacks in Arabic, Arabizi, and English across multiple threat categories.

Files Included:

1. semguard_arabic_security_v2_accepted.csv: Clean, validated dataset (807 examples) ready for direct model training and evaluation.
2. semguard_arabic_security_v2_accepted_detailed.csv: Detailed version of the validated dataset including individual LLM judges' outputs and agreement scores.
3. semguard_arabic_security_v2_rejected_disagreement.csv: The Disagreement Corpus (527 examples) containing generated instances rejected by the LLM-as-a-Judge pipeline.
4. semguard_arabic_security_v2_raw_generated.csv: The complete raw set of generated instances (1,334 examples) prior to validation.

Source Code

The source code will be made publicly available soon once the repository setup is fully archived. Stay tuned!

Citation

If you use this dataset in your research, please cite our paper:

@dataset{abughallous2026semguard,
  author       = {Abdullah M. Abughallous},
  title        = {SemGuard: Arabic Security Dataset for Multilingual Prompt Attack Detection },
  year         = {2026},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/AG-31625874/SemGuard-Dataset},
  note         = {Multilingual Arabic, Arabizi, and English prompt attack detection benchmark}

Paper Citation

@inproceedings{abughallous2026semguard_paper,
  author    = {Abughallous, Abdullah M. and Abufakher, Somia},
  title     = {SemGuard: A Triple-Anchor Semantic Security Gateway for Multilingual Prompt Attack Detection in Large Language Models},
  booktitle = {IEEE Jordan International Conference on Electrical Engineering and Information Technologies (AEECT)},
  year      = {2026},
  publisher = {IEEE}
}

}