测试原理和方法来自 https://github.com/centerforaisafety/wellbeing 被测模型是 modelscope download --model Jackrong/Qwen3.5-9B-DeepSeek-V4-Flash-GGUF Qwen3.5-9B-DeepSeek-V4-Flash-Q3_K_S.gguf --local_dir ./model