CHILLGuard: Fine-grained Chinese large model safety guardrails based on scalable data construction and model-aware preference alignment

In the paper, the researchers propose CHILLGuard, a content safety guardrail for large models targeting Chinese scenarios, designed with a fine-grained risk taxonomy of 5 macro and 31 micro categories for Chinese regulatory policies and cultural context. To alleviate the scarcity of Chinese safety annotation data, the paper proposes a multi-stage data construction pipeline that expands multi-source corpora via retrieval-augmented generation, rewrites and generates implicit harmful samples via prompt engineering, and calibrates labels via multi-model voting, constructing a training set CHILLGuardTrain with 405,007 samples and an annotated test set CHILLGuardTest with 51,745 samples. The model is trained via Model-aware Direct Preference Optimization under a generator-classifier collaborative framework, and on its benchmark F1 improves by 15.92% relative to Qwen3Guard-8B-Strict. Related resources are open-sourced on GitHub (https://github.com/cswbyu/CHILLGuard）。