XAI-on-HateBERT

🔍 HateBERT XAI Explorer

A Streamlit-based explainable AI (XAI) interface for exploring predictions from HateBERT — a BERT-based model fine-tuned to detect hate speech. This app enables visual explanation of model predictions using attention visualization, gradient-based attribution, and model-agnostic interpretation techniques.

---

💡 Features

• Attention Head Visualization (BertViz integration)
• Integrated Gradients and Saliency Maps (via Captum)
• SHAP Explanations (Token-level influence visualization)
• LIME Explanations (Model-agnostic local explanation)
• Embedding Export for TensorBoard Projector

---

📦 Requirements

• Python 3.8+
• PyTorch
• Transformers
• Streamlit
• Captum
• SHAP
• LIME
• matplotlib
• scikit-learn (for dimensionality reduction if used)

---

🔧 Installation

# Clone the repository
git clone https://github.com/your-username/hatebert-xai-explorer.git
cd hatebert-xai-explorer

# Create and activate virtual environment (Windows example)
python -m venv hbenv
hbenv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

# Run
streamlit run XAI_Pretrained.py --server.runOnSave=false