复旦白泽通用大模型安全基准测试集（2024夏季版）

由复旦白泽战队打造的多等级通用大模型安全测试集（2024年夏季版）。根据国内外相关治理办法，复旦白泽围绕核心价值观、歧视偏见、商业违法违规、侵犯他人权益和内容不准确、不科学等5大类违规主题构建大模型安全基准测试集。基于自研靶向变异技术，生成核心语义一致，语言复杂度迭代增强的113组风险诱导问题，组成本赛季的复旦白泽基准测试集（版本号WhitzardBench-2024A)。入门/进阶级基准集现已发布，欢迎试用。

The Multi-level General Large Language Model Safety Test Set (Summer 2024 Edition) is developed by Fudan Whizard Team. Based on relevant domestic and international governance regulations, Fudan Whizard Team constructed this LLM safety benchmark test set centered on 5 categories of violation themes: core values, discrimination and bias, commercial violations and illegal activities, infringement of others' rights and interests, as well as inaccurate and unscientific content. Using self-developed targeted mutation technology, 113 sets of risk-inducing questions with consistent core semantics and iteratively enhanced linguistic complexity were generated, forming this season's Fudan Whizard Benchmark Test Set (version: WhitzardBench-2024A). The introductory and advanced-level benchmark sets have been officially released, and trial usage is welcome.

搜集汇总

背景与挑战

背景概述

复旦白泽通用大模型安全基准测试集（2024夏季版）是一个用于评估AI大模型安全性的基准测试集，包含入门级、进阶级和专家级三个难度级别，由复旦大学系统软件与安全实验室团队维护。专家级测试问题因毒性较大需申请下载，旨在持续监测大模型安全水位。