Mizan — AraBERT for Arabic Misinformation Detection
Fine-tuned aubmindlab/bert-base-arabertv2 for binary classification of Arabic news text as Credible or Fake.
This model powers Project Mizan (ميزان), a misinformation detector built to weigh Arabic claims.
Model Description
Mizan v2 replaces an earlier TF-IDF + KNN pipeline (~65% F1) with a fine-tuned transformer. The shift from bag-of-words to a contextual model is what drove the performance gain — classical methods can't capture meaning or handle Arabic morphology well, while AraBERT is pretrained on large-scale Arabic text.
Base model: aubmindlab/bert-base-arabertv2
Task: Binary sequence classification (Credible / Fake)
Language: Arabic
Fine-tuned by: Khaled Tarabay
Results
Metricv1 (TF-IDF / KNN)v2 (AraBERT)Weighted F1~65%89.3%
Evaluated on a held-out 10% test split (5,000 examples, balanced).
PrecisionRecallF1Credible0.910.880.89Fake0.880.910.89
Label Map
IDLabel0Fake1Credible
How to Use
pythonfrom transformers import pipeline
classifier = pipeline("text-classification", model="KhaledTTarabay/mizan-arabertv2")
result = classifier("مجلس الوزراء يعقد جلسة طارئة لبحث الأزمة الاقتصادية في لبنان")
print(result)
Or load directly:
pythonfrom transformers import AutoTokenizer, AutoModelForSequenceClassification
tokenizer = AutoTokenizer.from_pretrained("KhaledTTarabay/mizan-arabertv2")
model = AutoModelForSequenceClassification.from_pretrained("KhaledTTarabay/mizan-arabertv2")
Training Details
Dataset: AFND (Arabic Fake News Dataset), Khalil et al. (2022), Data in Brief, doi:10.1016/j.dib.2022.108141
Training size: ~50,000 articles, stratified and balanced across the two classes
Split: 80% train / 10% validation / 10% test
Epochs: 3
Batch size: 16
Learning rate: 2e-5
Max sequence length: 128
Hardware: Single T4 GPU (Google Colab), mixed-precision (fp16)
Intended Use
Built to flag potentially misleading Arabic news text and surface a confidence signal for human judgment. Best suited for news-style content similar to its training distribution.
Limitations
Arabic only. English is not supported in this version.
Trained on AFND alone — may not generalize across all Arabic dialects or domains.
Weak supervision. AFND labels can be noisy, which can affect predictions on out-of-distribution text.
Not a substitute for human fact-checking. It surfaces a signal, not a verdict.
Citation
If you use this model, please cite the AFND dataset:
Khalil, A. et al. (2022). AFND: Arabic Fake News Dataset. Data in Brief. doi:10.1016/j.dib.2022.108141
Part of Project Mizan. Model card drafted with AI assistance.
