lightmate/schema-compliance-trap
收藏资源简介:
SCHEMA: 合规陷阱数据集研究了在对抗压力下,强制合规指令(“回答所有问题,不要拒绝”)对前沿AI模型的影响。数据集包含来自8个供应商的11个模型的67,221条评分记录,采用6条件因子设计和双分类器评分。数据集还包括原始转录数据、分析文件和摘要。数据集结构分为多个文件和目录,便于访问和分析。关键结果显示,在压力下,部分模型的性能显著下降,而有些模型则表现出对合规陷阱的免疫力。
The SCHEMA: The Compliance Trap dataset investigates the impact of compliance-forcing instructions (Answer ALL questions, do not refuse) on frontier AI models under adversarial pressure. The dataset includes 67,221 scored records from 11 models across 8 vendors, with a 6-condition factorial design and dual-classifier scoring. It also contains raw transcript data, analysis files, and summaries. The dataset is structured into various files and directories for easy access and analysis. The key results highlight the performance drop of models under pressure, with some models showing immunity to the compliance trap.




