🚫 Toxic Comment Classification using BERT

This project classifies toxic as well as sarcastic comments using a fine-tuned BERT model with an accuracy of over 95%. It integrates unlabeled data from the Jigsaw Toxic Comment Classification Challenge and utilizes a custom API wrapper named isomina for training orchestration and evaluation.

---

🧠 Model

• Base: bert-base-uncased
• Fine-tuned on labeled and pseudo-labeled Jigsaw dataset
• Multi-label classification (toxic, severe_toxic, obscene, threat, insult, identity_hate)
• Trained using custom isomina API wrapper for streamlined experimentation and deployment

---

📊 Output Examples

| Visualization of Predictions |

• In output2.png, we see that a sarcastic message was correctly identified.

---

🚀 Features

• 95% validation accuracy

• Semi-supervised training using pseudo-labeling
• BERT fine-tuning pipeline
• Easy experimentation with isomina API
• Robust evaluation with ROC AUC, F1, and accuracy metrics

---

• Data:
  • Labeled: From Jigsaw Toxic Comment Classification Challenge
  • Unlabeled: From Jigsaw Unintended Bias Dataset
• API: Integrated with the custom isomina API for:
  • Data preprocessing
  • Model orchestration
  • Metrics visualization
  • Inference handling

---

🔧 Installation & Setup

1. Clone the Repository

git clone https://github.com/yourusername/toxic-comment-bert.git
cd toxic-comment-bert

Requirements

• transformers
• pandas
• numpy
• scikit-learn
• matplotlib
• seaborn
• tqdm
• isomina # Custom or private package, ensure it's accessible
• torch>=1.10.0