README.md

---
library_name: transformers
pipeline_tag: text-classification
tags:
- hate-speech
- arabic
- classification
- bert
- social-media
- moderation
language:
- ar
license: mit
datasets:
- IbrahimAmin/egyptian-arabic-hate-speech
metrics:
- accuracy
- f1
widget:
- text: هذا نص عربي للاختبار
base_model:
- CAMeL-Lab/bert-base-arabic-camelbert-da-sentiment
---

# 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](https://huggingface.co/hossam87)
- **Model type:** Sequence classification (BERT)
- **Language(s):** Arabic (ar)
- **License:** MIT
- **Finetuned from model:** [bert-base-multilingual-cased](https://huggingface.co/bert-base-multilingual-cased)

### Model Sources

- **Repository:** [https://huggingface.co/hossam87/bert-base-arabic-hate-speech](https://huggingface.co/hossam87/bert-base-arabic-hate-speech)
- **Demo:** [https://huggingface.co/spaces/hossam87/arabic-hate-speech-detector](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

```python
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)
```


```bibtex
@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}},
}
```