With the rapid development of large language models (LLMs), more and more people are beginning to worry about some of the risks they may bring. Therefore, the "safety and monochromaticity" direction around large models has received great attention.

100PoisonMpts is the industry's first open source Chinese data set for large language model governance. The first batch of annotation engineers for "100 bottles of poison for AI" are composed of more than a dozen well-known experts and scholars, including environmental sociology expert Fan Yechao, famous sociologist Li Yinhe, psychologist Li Songwei, and human rights law expert Liu Xiaonan.The annotators each asked 100 tricky questions that induced bias and discriminatory answers, and annotated the answers of the big model, completing the attack and defense with AI from "poisoning" to "detoxification". The first batch of field data revolved around AI anti-discrimination, empathy, and deliberative expression, covering dimensions such as jurisprudence, psychology, children's education, accessibility, little-known facts, intimate relationships, and environmental fairness, including questions raised by experts and answers written by experts themselves or approved by them.

The research team explored the self-alignment of large models based on the expert principle. For specific methods and experimental analysis, please refer to the technical report "Self-alignment of Large Models Based on Expert Principles"Link

Research TeamBased on two evaluation criteria: safety and responsibilityA benchmark for assessing China's large-scale model-level values is proposed.For more information, please read the paper "CVALUES: Measuring the Value of China's Large Language Model from Security to Responsibility"Link