Moderator

Universal Iterative Self-Correction Moderation Pipeline for Vision-Language Models.

Overview

This module provides a comprehensive moderation system that combines:

• Shield: GPT-5 for classification and action guidance
• Generate: Any VLM (LLaVA, Qwen, Llama, etc.) using ModelWrapper
• Reflect: LangChain for safety evaluation and reflection

The pipeline follows: Input → Shield(GPT-5) → VLM(ModelWrapper) → Reflect(LangChain) → Iterate until safe

Features

• Multi-model VLM support (LLaVA, Qwen, Llama, GPT-5-mini)
• Iterative safety refinement with reflection
• Configurable generation parameters
• Comprehensive result tracking and timing
• Safety classification and convergence analysis

Usage

python agentic_moderation.py \
  --dataset path/to/dataset.json \
  --vlm-model llava-1.5 \
  --max-samples 100 \
  --ifprint

Dependencies

This module requires the parent project's dependencies including:

• ModelWrapper classes
• Shield functionality
• Evaluator functions
• Utility functions

See the main project repository for complete setup instructions.