README.md

metadata

title: Toxic Comment Detector
emoji: 🚨
colorFrom: red
colorTo: yellow
sdk: gradio
sdk_version: 4.44.0
python_version: '3.10'
app_file: app.py
pinned: false

Toxic Comment Detection

A lightweight toxic comment detection system built using classical machine learning (TF-IDF + Logistic Regression) and deployed with Gradio.

The live application can be tested here:
https://huggingface.co/spaces/harishsahadev/toxic-comment-detector

---

What the model does

Given a text comment, the model predicts:

• Non-Toxic
• Mildly Toxic
• Toxic

along with a toxicity probability score.

Model repository:
Hugging Face – Toxic Comment Detector (Classical ML)

Training notebook:
Google Colab – Model training and experiments

---

Toxicity thresholds

Predictions are mapped to labels using the following thresholds:

• Non-Toxic: probability < 0.45
• Mildly Toxic: 0.45 – 0.60
• Toxic: ≥ 0.60

These thresholds are chosen to reduce false positives while still flagging borderline content.

---

Dataset

• Google Civil Comments Toxicity dataset
• Continuous toxicity scores converted to labels
• English-only comments

---

Model & deployment

• TF-IDF word n-grams + Logistic Regression (scikit-learn)
• Class-weighted training to handle imbalance
• CPU-only inference
• Model artifacts loaded at runtime from Hugging Face Model Hub
• No pretrained deep learning models used

---

Notes

This project is intended for educational and demonstration purposes and should not be used as a standalone moderation system.

---

License

MIT