Dual-Stage Toxic Moderation App

How to run locally

Reproduce the Streamlit app locally using the project requirements and a local copy of the fine-tuned model.

pip install -r requirements.txt
# copy .env.example to .env and add your OpenRouter API key
# place model weights under the path used by the app (see README)
streamlit run app_streamlit.py

Model files used in Week 2 are expected at: C:/Users/NightPrince/OneDrive/Desktop/Cellula-Internship/Week1/peft-distilbert-toxic-classifier/last-checkpoint/ — update paths in the app if needed.

• Modular dual-stage moderation pipeline: Llama Guard (hard filter) followed by DistilBERT+LoRA (9-class classifier).
• Multi-modal support via BLIP image captioning to handle images through the same text pipeline.
• Production-ready Streamlit UI with clear feedback, class probabilities, and error handling.
• Addressed class imbalance using SMOTE, oversampling, and class weights; documented experiments in reports.
• Deployed a public demo on Hugging Face Spaces for evaluation and sharing.

Week 2 File Structure (excerpt)

Week2/
├── app_streamlit.py         # Main Streamlit app
├── pipeline/                # Modular pipeline
│   ├── blip_caption.py
│   ├── llama_guard.py
│   └── toxic_classifier.py
├── requirements.txt
├── .env.example
├── README.md
└── internship_week2_report.html
                

1. Dual-Stage Architecture Design: Designed a modular dual-stage pipeline with Stage 1 (LlamaGuard hard filter) and Stage 2 (DistilBERT+LoRA fine-grained classification) for comprehensive content moderation.
2. Stage 1 - LlamaGuard Integration: Integrated LlamaGuard via OpenRouter API for instant hard filtering of legally or ethically unsafe content, ensuring only 'safe' or 'unsafe' responses for maximum reliability.
3. Stage 2 - DistilBERT+LoRA Implementation: Deployed the fine-tuned DistilBERT model with PEFT-LoRA for nuanced 9-class toxic content classification, addressing class imbalance through SMOTE and weighted loss functions.
4. Multi-Modal BLIP Integration: Implemented BLIP model for image captioning, enabling visual content moderation by converting images to text and processing through the same dual-stage pipeline.
5. Modular Pipeline Development: Built a clean, modular architecture with separate components for BLIP captioning, LlamaGuard filtering, and toxic classification, ensuring maintainability and extensibility.
6. Production-Ready Streamlit App: Created a robust Streamlit application with comprehensive error handling, real-time processing, professional UI, and support for both text and image inputs.

Key Achievements

• Successfully implemented a production-ready dual-stage moderation system with real-time processing
• Created a modular, extensible pipeline architecture with clean separation of concerns
• Integrated multiple state-of-the-art models (LlamaGuard, DistilBERT+LoRA, BLIP) into a unified system
• Built a comprehensive Streamlit application with professional UI and robust error handling
• Deployed live demo on Hugging Face Spaces for public access and testing

• Dual-Stage Pipeline Integration: Successfully integrated LlamaGuard (Stage 1) and DistilBERT+LoRA (Stage 2) by creating a unified interface and handling different API formats and model outputs.
• Multi-Modal Processing: Implemented BLIP image captioning and developed algorithms to effectively combine text and image analysis results for comprehensive safety evaluation.
• Modular Architecture Design: Designed a clean, modular pipeline with separate components for BLIP captioning, LlamaGuard filtering, and toxic classification, ensuring maintainability and extensibility.
• Production Deployment: Optimized model loading, inference pipelines, and error handling to achieve real-time performance while maintaining accuracy for production deployment on Hugging Face Spaces.