saiheinthuyasoe/offensive-language-detector-gradio-interface

🛡️ Offensive Language Detection Model

This Hugging Face Space hosts an offensive language detection model built with DistilBERT. The model can classify text into three categories:

• —hate_speech: Content targeting specific groups based on identity (race, religion, gender, etc.)
• —offensive_language: General offensive content, insults, or profanity
• —neither: Non-offensive content

🚀 Features

• —Real-time Analysis: Instant classification of text input
• —Confidence Scores: Get probability scores for all categories
• —Interactive Interface: Easy-to-use Gradio web interface
• —Example Texts: Pre-loaded examples to test the model
• —Detailed Results: Complete breakdown of predictions

🔧 Model Details

• —Base Model: DistilBERT (distilled version of BERT)
• —Task: Multi-class text classification
• —Classes: 3 (hatespeech, offensivelanguage, neither)
• —Framework: Transformers + PyTorch

📊 Usage

1. 1.Enter your text in the input box
2. 2.Click "Analyze Text" or press Enter
3. 3.View the prediction results:
4. 4.Predicted category
5. 5.Confidence score
6. 6.Whether the text is considered offensive
7. 7.Probability distribution across all classes

⚠️ Important Notes

• —This model is for research and educational purposes
• —Always review automated decisions in production systems
• —The model may have biases based on its training data
• —Consider human oversight for critical content moderation tasks

🛠️ Technical Implementation

The model uses:

• —Tokenizer: DistilBERT tokenizer with 512 max tokens
• —Architecture: DistilBERT for sequence classification
• —Inference: PyTorch with softmax for probability distribution
• —Interface: Gradio for interactive web UI

📝 Example Classifications

• —"I love spending time with family" → neither (high confidence)
• —"This movie is terrible" → offensive_language (medium confidence)
• —"You're an idiot" → offensive_language (high confidence)
• —Identity-based attacks → hate_speech (varies by content)

🤝 Contributing

This model can be improved through:

• —Better training data with balanced examples
• —Fine-tuning on domain-specific content
• —Adjusting classification thresholds
• —Adding more nuanced categories

📄 License

MIT License - Feel free to use and modify for your projects.

Built with ❤️ using Hugging Face Transformers and Gradio