Team Ai
Apppublic

YaekobB/Toxic-Comment-Classification

sourceHugging Facemitupdated 5mo agoView on Hugging Face
0likes
App README

🧠 Toxic Comment Classification — Explainable Multi-Label NLP Model

<p align="center"> <img src="banner.png" alt="Toxic Comment Classification Banner" width="100%"> </p>

<p align="center"> <b>DistilBERT-based multi-label classifier for detecting toxic online comments with explainability powered by Captum Integrated Gradients (IG).</b> </p>

🚀 Overview

This project presents an explainable AI system for identifying toxic comments in text, built using a fine-tuned Transformer model (DistilBERT). It performs multi-label classification across six toxicity categories while offering token-level explanations for each prediction.

🧩 Labels

  • —toxic
  • —severe_toxic
  • —obscene
  • —threat
  • —insult
  • —identity_hate

🎯 Objectives

  • —Fine-tune DistilBERT for robust multi-label toxicity detection
  • —Enhance interpretability using Captum Integrated Gradients
  • —Deploy a real-time, user-friendly Gradio interface

🧪 How to Use the Demo

  1. 1.Type or paste any comment in the text box
  2. 2.Click “Classify” to view per-label probabilities and predictions
  3. 3.Open the “Explain” tab → select a target label
  4. 4.Generate a heatmap showing which words support (red) or oppose (blue) the decision

🧠 Example Inputs

ExampleExpected Labels
“You are a complete idiot.”toxic / insult
“I will kill you tomorrow.”threat / toxic
“Thanks for your help today!”non-toxic
“Go away, you people don’t belong here.”identity_hate / insult

⚙️ Technical Stack

ComponentTechnology
Language ModelDistilBERT (distilbert-base-uncased)
FrameworksPyTorch • Transformers • Gradio
ExplainabilityCaptum (Integrated Gradients)
TrainingStratified splits • Early Stopping • Regularization
VisualizationGradio UI + Captum HTML heatmaps
DeploymentHugging Face Spaces

📂 Project Structure

.
├── app.py                # Gradio app entry point
├── requirements.txt      # Runtime dependencies
├── artifacts/
│   ├── best/             # Fine-tuned model weights + tokenizer
│   └── thresholds.json   # Tuned thresholds for each label
└── README.md             # (this file)

📊 Model Training Summary

  • —Dataset: Jigsaw Toxic Comment Classification Challenge
  • —Tokenization: DistilBERT (max length = 256)
  • —Loss: Binary Cross-Entropy with Logits (BCEWithLogitsLoss)
  • —Optimizer: AdamW (learning rate = 2e-5, weight decay = 0.02)
  • —Regularization: Dropout (head=0.5, encoder=0.2)
  • —Evaluation Metrics: Macro F1 • Precision • Recall • AUC
  • —Explainability: Captum Layer Integrated Gradients (LIG)

🖥️ Live Demo

🚀 Try the interactive demo on Hugging Face Spaces: 🔗 [yaekobB / Toxic-Comment-Classification](https://huggingface.co/spaces/yaekobB/Toxic-Comment-Classification)

🧰 Dependencies

txt
transformers>=4.41.0
torch>=2.2.0
safetensors>=0.4.2
gradio>=4.20.0
captum>=0.7.0
pandas>=2.0.0
numpy>=1.24.0

🪪 License

This project is licensed under the MIT License. You are free to use, modify, and distribute this work with attribution.

<p align="center"> <i>“Building safer and explainable AI for online interactions.”</i> </p>