🚫 NSFW Text Detection using BigBird + Hugging Face Deployment

This repository contains a complete pipeline for fine-tuning Google's BigBird transformer for NSFW (Not Safe For Work) text classification and deploying it as a live API on Hugging Face Spaces using Gradio. The solution is optimized for identifying inappropriate textual content in long-form documents — useful for child protection systems, content moderation, and safe browsing applications.

---

📌 Features

• ✅ Fine-tuned google/bigbird-roberta-base model for multi-label NSFW text classification.
• ✅ Supports long text inputs (up to 4096 tokens).
• ✅ Lightweight inference API with Gradio.
• ✅ Hosted on Hugging Face Spaces for public access.
• ✅ Google Colab training notebook included.
• ✅ Easily customizable for other text classification tasks.

---

🧠 Model Overview

• Architecture: BigBird-RoBERTa-base
• Task: Binary classification (Safe = 0, NSFW = 1)
• Input: Raw text paragraphs
• Output: 0 (Safe), 1 (Not Safe)

The model is trained on a labeled dataset (NSFW1.csv) consisting of text segments with corresponding binary labels indicating safety.

---

🛠️ Installation

To run the API locally:

git clone https://github.com/rusiru-erandaka/Bigbird_huggingface_deploy.git
cd Bigbird_huggingface_deploy
pip install -r requirements.txt
python app.py

---

📁 Project Structure

Bigbird_huggingface_deploy/

• Bigbird2_.ipynb # Training notebook (Google Colab)
• app.py # Gradio API app
• onfig.json # Model config
• tokenizer_config.json # Tokenizer config
• pecial_tokens_map.json # Tokenizer special tokens
• spiece.model # SentencePiece model
• requirements.txt # Python dependencies
• LICENSE

---

🚀 Deployment on Hugging Face

• Upload your fine-tuned model and tokenizer files to the Hugging Face Model Hub.
• Deploy app.py on Hugging Face Spaces using Gradio as the interface.
• Ensure that the model_id in app.py matches your model repository name, e.g., "Rerandaka/Cild_safety_bigbird".

---

🔍 Example Usage

from transformers import pipeline
classifier = pipeline("text-classification", model="Rerandaka/Cild_safety_bigbird")
classifier("This is an inappropriate message involving violence.")
# Output: [{'label': 'LABEL_1', 'score': 0.98}]

---

📊 Training Summary

• Optimizer: AdamW
• Epochs: 3–5
• Evaluation Metrics: Accuracy, Precision, Recall
• Dataset Split: 80% Train / 20% Test
• Platform: Google Colab with GPU

---

📎 Dependencies

transformers
torch
gradio

---

📜 License

This repository is licensed under the MIT License.

---

🙋‍♂️ Acknowledgements

• Hugging Face 🤗 for Transformers and Spaces
• Google for BigBird-RoBERTa
• Gradio for interactive UI
• Rusiru Erandaka for fine-tuning and deployment