This project implements an explainable multi-label toxic comment classification system using the DistilBERT transformer model.
It is based on the Jigsaw Toxic Comment Classification Challenge dataset and predicts six forms of online toxicity while providing interpretable visual explanations using Captum Integrated Gradients (IG).

You can try the interactive web app on Hugging Face Spaces:
👉 https://huggingface.co/spaces/YaekobB/Toxic-Comment-Classification

✅ Fine-tuned DistilBERT model on 6 toxicity labels
✅ Handles multi-label text classification (comments can belong to multiple categories)
✅ Explainable AI (XAI) with Captum to visualize token-level attributions
✅ Gradio-based UI for real-time text classification and interpretation
✅ Clean, reproducible end-to-end Kaggle notebook

Toxic-Comment-Classification/
│
├── README.md                                   # Project overview and documentation
├── requirements.txt                            # Python dependencies
├── Toxic_Comment_Classification_full.ipynb     # End-to-end Kaggle notebook
├── demo/
│   └── app.py                                  # Gradio UI for local demo or Spaces
│

git clone https://github.com/yaekobB/Toxic-Comment-Classification.git
cd Toxic-Comment-Classification

pip install -r requirements.txt

python demo/app.py

Access the local app at → http://127.0.0.1:7860

The fine-tuned model is available on Hugging Face Hub:

👉 https://huggingface.co/spaces/YaekobB/Toxic-Comment-Classification

To use the model locally, download the weights and place them inside:

artifacts/best/

Explainability Heatmap

Color legend:
🔴 Red — Words that increase the toxicity score
🔵 Blue — Words that decrease the toxicity score

Each word’s intensity reflects its contribution to the final prediction.

MIT License © 2025 You are free to use, modify, and distribute this project with proper attribution.