WalidAlsafadi/arabert-fake-news-detector

AraBERT Fake News Detector

Model Overview

This model is a fine-tuned version of AraBERT (aubmindlab/bert-base-arabertv02) for detecting fake news in Arabic news articles, with a focus on Palestinian news sources.

Intended Use

• Task: Binary classification (Real vs. Fake news) for Arabic text.
• Audience: Journalists, researchers, and the general public interested in verifying Arabic news content.
• Input: Arabic news article (max 512 tokens).
• Output: "Real" or "Fake" label with confidence scores.

Dataset

• News articles scraped from Palestinian news sources.
• Cleaned and split into training, validation, and test sets.
• Test set size: 2750 samples.

Performance

• Accuracy: 96.22%
• F1-score (weighted): 96.22%
• F1-score (macro): 96.21%
• AUC: 99.57%
• Inference time: ~200ms (CPU)

Confusion Matrix

[[1251   32]
 [  72 1395]]

Classification Report

Class	Precision	Recall	F1-score	Support
Real	0.9456	0.9751	0.9601	1283
Fake	0.9776	0.9509	0.9641	1467

Limitations & Biases

• Trained on Palestinian news; performance may vary on other Arabic dialects or regions.
• Potential bias from dataset sources and labeling.
• Not suitable for non-news or non-Arabic text.

Ethical Considerations

• Use responsibly; predictions are not a substitute for human judgment.
• False positives/negatives may occur.

Citation

If you use this model, please cite:

@misc{haqiqa2025,
  author = {Walid Alsafadi},
  title = {Haqiqa - Arabic Fake News Detector},
  year = {2025},
  url = {https://github.com/WalidAlsafadi/Haqiqa-Arabic-Fake-News-Detector}
}

License

Licensed under Apache 2.0 - see LICENSE for details.