Modelpublic

AmrMohamed21/arabert-fake-news

sourceHugging Faceapache-2.0updated 4mo agoView on Hugging Face

0likes12downloads

Model Card

🕵️ Arabic News Credibility Analyzer

A fine-tuned AraBERT v2 model for detecting fake news and credibility classification in Arabic text. Trained on 6,267 Arabic news articles, achieving 98.8% accuracy on the test set.

📋 Model Details

Property	Details
Base Model	`aubmindlab/bert-base-arabertv02`
Task	Binary Text Classification (Real vs Fake)
Language	Arabic (Modern Standard + Dialectal)
Training Samples	6,267 Arabic news articles
Test Accuracy	98.8%
Framework	PyTorch + HuggingFace Transformers

🚀 Quick Start

python

from transformers import pipeline

# Load model directly
classifier = pipeline(
    "text-classification",
    model="AmrMohamed21/arabert-fake-news"
)

text = "أعلنت الحكومة اليوم عن خطة اقتصادية جديدة لدعم القطاع الصناعي"
result = classifier(text)
print(result)
# [{'label': 'REAL', 'score': 0.997}]

🔧 Load Model Manually

python

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

model_name = "AmrMohamed21/arabert-fake-news"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

def predict(text: str) -> dict:
    inputs = tokenizer(
        text,
        return_tensors="pt",
        truncation=True,
        max_length=512,
        padding=True
    )
    
    with torch.no_grad():
        outputs = model(**inputs)
        probabilities = torch.nn.functional.softmax(outputs.logits, dim=-1)
    
    labels = ["FAKE", "REAL"]
    predicted_class = probabilities.argmax().item()
    
    return {
        "label": labels[predicted_class],
        "confidence": probabilities[0][predicted_class].item()
    }

# Example
text = "مصدر مجهول يدعي سقوط الحكومة غداً بشكل مفاجئ"
print(predict(text))

📊 Training & Evaluation

Training Configuration

python

training_args = TrainingArguments(
    output_dir="./arabert-fake-news",
    num_train_epochs=5,
    per_device_train_batch_size=16,
    per_device_eval_batch_size=16,
    learning_rate=2e-5,          # Carefully tuned to avoid catastrophic forgetting
    warmup_steps=500,
    weight_decay=0.01,
    evaluation_strategy="epoch",
    save_strategy="epoch",
    load_best_model_at_end=True,
)

Results

Metric	Score
Accuracy	98.8%
F1 Score	98.7%
Precision	98.9%
Recall	98.6%

Confusion Matrix Highlights

—✅ Very low false positive rate (misclassifying real news as fake)
—✅ Strong performance on both Modern Standard Arabic and news-style text

🏗️ Architecture

Input Text (Arabic)
       ↓
[AraBERT v2 Tokenizer]   ← Handles Arabic morphology
       ↓
[BERT Encoder × 12]      ← Pre-trained Arabic language understanding
       ↓
[CLS Token Representation]
       ↓
[Dropout(0.1)]
       ↓
[Linear(768 → 2)]        ← Classification head
       ↓
[Softmax]
       ↓
Output: REAL / FAKE + confidence score

Why AraBERT v2?

—Pre-trained specifically on Arabic text (24GB Arabic corpus)
—Handles Arabic morphological complexity (root-based structure)
—Understands Arabic news-specific vocabulary

📁 Dataset

The model was trained on 6,267 Arabic news articles collected from:

—Arabic news websites (real articles)
—Known disinformation sources (fake articles)

Split	Samples
Train	~5,013 (80%)
Validation	~627 (10%)
Test	~627 (10%)

⚠️ Note: Dataset contains Arabic news text from various domains including politics, economy, and social news.

⚠️ Limitations & Bias

—Trained primarily on Modern Standard Arabic — may show reduced accuracy on heavy dialects
—Performance may vary on domain-specific content (medical, legal, etc.)
—News landscape evolves — model reflects training data from its collection period
—Should not be used as the sole decision-maker for content moderation

🔗 Related Links

Resource	Link
🐙 GitHub Repository	Arabic-News-Credibility-Analyzer
👤 HuggingFace Profile	@AmrMohamed21
📦 Base Model	aubmindlab/bert-base-arabertv02

📄 Citation

If you use this model in your research or project, please cite:

bibtex

@misc{amrmohamed21-arabert-fake-news-2024,
  author       = {Amr Mohamed},
  title        = {Arabic News Credibility Analyzer: Fine-tuned AraBERT v2 for Fake News Detection},
  year         = {2024},
  publisher    = {HuggingFace},
  journal      = {HuggingFace Model Hub},
  howpublished = {\url{https://huggingface.co/AmrMohamed21/arabert-fake-news}}
}

👨‍💻 About the Author

Built by Amr Mohamed — ML/AI Engineer specializing in Arabic NLP and production-grade AI systems.

—🤗 HuggingFace: @AmrMohamed21
—🐙 GitHub: @Amr-Mo-ali

Made with ❤️ for Arabic NLP