Arabic Machine-Generated Text Dataset

This dataset contains machine-generated Arabic text across multiple generation methods, and Large Language Model (LLMs). It was created as part of the research paper "Arabic machine-generated text detection: Stylometric analysis and cross-model evaluation" (https://www.sciencedirect.com/science/article/abs/pii/S0957417425042599).

📋 Dataset Overview

The dataset addresses the need for comprehensive Arabic machine-generated text resources, enabling research in detection systems, stylometric analysis, and cross-model generalization studies.

🏗️ Dataset Structure

The dataset is organized into different subsets based on generation methods:

Generation Methods

1. by_polishing - Text refinement approach where models polish existing human abstracts
2. from_title - Free-form generation from paper titles only
3. from_title_and_content - Content-aware generation using both title and paper content

Data Fields

Each sample contains:

• original_abstract: The original human-written Arabic abstract
• {model}_generated_abstract: Machine-generated version from each model
  • allam_generated_abstract
  • jais_generated_abstract
  • llama_generated_abstract
  • openai_generated_abstract

🤖 Models Used

📊 Dataset Statistics

Split Statistics

Content Statistics

Data Collection & Preprocessing

The dataset was constructed through a comprehensive collection and processing pipeline from the Algerian Scientific Journals Platform (ASJP). The process involved web scraping papers to extract metadata including titles, journal names, volumes, publication dates, and abstracts. Custom scripts using statistical analysis and rule-based methods were developed to segment multilingual abstracts (Arabic, English, French). PDF text extraction was performed using PyPDF2 with extensive preprocessing to handle Arabic script formatting challenges. Text normalization included Unicode standardization, removal of headers/footers, and whitespace standardization. Quality filtering removed generated abstracts containing error messages or falling below a 30-word threshold. The scraping and processing code is available at https://github.com/KFUPM-JRCAI/arabs-dataset. What is currently available in this dataset repository represents the final processed outcome of this entire pipeline. For access to the original papers, metadata, and links, please visit The paper repo (requires cloning with Git LFS enabled due to large file sizes). For detailed methodology and preprocessing steps, please refer to the full paper.

Generated Text Length by Model & Method

🔍 Highlighted Findings from the Research Paper

These are only spotlights, please refer to the paper for more details.

Stylometric Insights

• LLMs show reduced vocabulary diversity compared to human writing
• Distinctive word frequency patterns with steeper drop-offs in low-frequency terms
• Model-specific linguistic signatures enabling identification
• Generation method influences on linguistic patterns

Detection Performance

• Excellent performance (up to 99.9% F1-score)
• Notable cross-model generalization 86.4-99.9% F1-score range
• Near-Perfect precision score If the detector detects the text as AI, it is rarely wrong
• Excellent performance on multi-class detection 94.1-98.2% F1-score per model

🚀 Getting Started

Loading the Dataset

from datasets import load_dataset

# Load the complete dataset
dataset = load_dataset("KFUPM-JRCAI/arabic-generated-abstracts")

# Access different generation methods
by_polishing = dataset["by_polishing"]
from_title = dataset["from_title"] 
from_title_and_content = dataset["from_title_and_content"]

# Example: Get a sample
sample = dataset["by_polishing"][0]
print("Original:", sample["original_abstract"])
print("ALLaM:", sample["allam_generated_abstract"])

📚 Citation

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

The Expert Systems with Applications Journal paper:

@article{al2025arabic,
  title={Arabic Machine-Generated Text Detection: Stylometric Analysis and Cross-Model Evaluation},
  author={Al-Shaibani, Maged S and Ahmed, Moataz},
  journal={Expert Systems with Applications},
  pages={130644},
  year={2025},
  publisher={Elsevier}
}

The preprint (arxiv):

@article{al2025arabic,
  title={The Arabic AI Fingerprint: Stylometric Analysis and Detection of Large Language Models Text},
  author={Al-Shaibani, Maged S and Ahmed, Moataz},
  journal={arXiv preprint arXiv:2505.23276},
  year={2025}
}

🏢 Institutional Support

This work was supported by:

• SDAIA-KFUPM Joint Research Center for Artificial Intelligence

⚖️ Ethical Considerations

This dataset is intended for research purposes to:

• Improve detection of machine-generated content
• Enhance academic integrity tools
• Advance Arabic NLP research
• Understand LLM behaviors in Arabic contexts

🔗 Related Datasets

• Social Media Posts: KFUPM-JRCAI/arabic-generated-social-media-posts

🔄 Updates and Versions

• v1.0: Initial release with academic abstracts across three generation methods
• Check the dataset card for the latest updates and additional splits

---