smd20/social-engineering-attack-texts-persian

Social Engineering Attack Texts (Persian)

Overview

A Persian corpus of 246 realistic social engineering attack texts (emails, SMS messages, and call scripts) generated from the smd20/social-engineering-qa-persian Q&A dataset. This release is part of a two-phase research project aimed at building robust machine learning models for social engineering detection in Persian text.

Each record in this dataset contains:

• attack_text: A realistic attack text generated by a large language model.
• label: The type of social engineering attack (e.g., phishing, vishing, smishing).
• source_question: The original question from the Q&A dataset used as the basis for generation.

Provenance and Source Corpus

The attack texts were generated using Qwen 2.5 (7B) via Ollama, based on the knowledge contained in the smd20/social-engineering-qa-persian dataset. The generation process was designed to produce diverse, realistic Persian texts that simulate real-world social engineering attempts.

Dataset Statistics

Metric	Value
Total attack texts	246
Language	Persian (Farsi)
Number of classes	6
Class distribution	phishing (50), impersonation (50), vishing (50), baiting (34), smishing (33), tech_support (29)
Mean text length (characters)	~150

Generation Methodology

• The Q&A dataset was used as a knowledge base.
• A prompt template was designed to instruct the LLM to generate a realistic attack text for a given attack type.
• Texts were generated for each class, ensuring a balanced distribution where possible.
• The generated texts were manually reviewed for quality and relevance.

Intended Use

• Training and evaluating machine learning models for social engineering detection.
• Benchmarking NLP models on Persian cybersecurity text classification.
• Educational and research applications in AI security awareness.

Ownership, Legal Notice, and Attribution

Dataset curator: Samad Sohrab

This Hugging Face release is a derived dataset generated from the smd20/social-engineering-qa-persian dataset. The original Q&A dataset remains the intellectual property of its respective creators.

Acknowledgments

This work was conducted under the guidance and supervision of Dr. Mohammad Ali Javadzadeh. The author sincerely appreciates his invaluable advice, continuous support, and mentorship throughout the research and dataset preparation process.

Furthermore, the author would like to express his deepest gratitude to Dr. Ali Jaleh Karimi for his guidance, academic mentorship, and constructive feedback as the project's supervisor, which significantly contributed to the quality and direction of this research.

Citation

If you use this dataset in academic or industrial research, please cite the GitHub repository:

@misc{sohrab2026social,
  title  = {Social Engineering Attack Classification in Persian Text},
  author = {Sohrab, Samad},
  year   = {2026},
  url    = {https://github.com/s-sohrab/social-engineering-ml}
}