CHILLGuard: Towards Fine-Grained Chinese LLM Safety Guardrail with Scalable Data Construction and Model-aware Preference Alignment
CHILLGuard is a new safety system for Chinese LLMs that addresses cultural and regulatory risks by using a detailed harm taxonomy. It combines a scalable data pipeline with Model-aware Direct Preference Optimization (MDPO) to train a robust guardrail. This method significantly improves detection accuracy, achieving state-of-the-art performance across different model sizes.