Model Card for hossam87/bert-base-arabic-hate-speech

A fine-tuned BERT model to classify Arabic text into: Neutral, Offensive, Sexism, Religious Discrimination, or Racism.

---

Model Details

Model Description

This model is based on bert-base-multilingual-cased and fine-tuned on an Arabic social media dataset for hate speech detection.
It classifies Arabic text into one of five categories: Neutral, Offensive, Sexism, Religious Discrimination, or Racism.
Intended uses include moderation, analytics, and academic research.

• Developed by: hossam87
• Model type: Sequence classification (BERT)
• Language(s): Arabic (ar)
• License: MIT
• Finetuned from model: bert-base-multilingual-cased

Model Sources

• Repository: https://huggingface.co/hossam87/bert-base-arabic-hate-speech
• Demo: https://huggingface.co/spaces/hossam87/arabic-hate-speech-detector

Training Details

Training Data

The model was fine-tuned on a labeled dataset of Arabic social media posts, manually annotated for the five target categories.

Training Procedure

• Precision: Mixed precision (fp16)
• Epochs: 4 (best model at epoch 3)
• Batch size: 32
• Learning rate: 3e-5
• Optimizer: AdamW
• Hardware: 2 x NVIDIA T4 GPUs (Kaggle)

---

Evaluation

Metrics

Metric	Score
Accuracy	0.944
F1 Macro	0.946

Uses

Direct Use

• Content moderation for Arabic social media, forums, and chats.
• Analytics and research into hate speech patterns in Arabic.
• Educational and academic projects.

Out-of-Scope Use

• Automated moderation without human oversight in sensitive or legal contexts.
• Use on languages other than Arabic.
• General text classification tasks outside hate speech detection.

Bias, Risks, and Limitations

The model may misclassify:

• Sarcasm, slang, or context-dependent expressions.
• Formal written Arabic, since trained on social media content.
• Domain-specific or emerging hate speech not represented in the training data.

Recommendations

Always keep a human-in-the-loop for sensitive moderation tasks. Use responsibly and be transparent about automation.

How to Get Started with the Model

from transformers import AutoTokenizer, AutoModelForSequenceClassification, pipeline

model_id = "hossam87/bert-base-arabic-hate-speech"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(model_id)

classifier = pipeline("text-classification", model=model, tokenizer=tokenizer)

text = "هذا نص عربي للاختبار"
result = classifier(text)
print(result)

@misc{hossam87_2025_arabichate,
  title = {BERT-base Arabic Hate Speech Detector},
  author = {Hossam87},
  year = {2025},
  howpublished = {\url{https://huggingface.co/hossam87/bert-base-arabic-hate-speech}},
}