LLM Physical Safety Benchmark Datasets
收藏arXiv2025-09-30 收录
下载链接：
https://huggingface.co/datasets/TrustSafeAI/llm_physical_safety_benchmark
下载链接
链接失效反馈官方服务：
资源简介：
该数据集是一个精心策划的集合，包含超过400种不同的指令，旨在评估大型语言模型在无人机控制场景中的物理安全性和实用性。该数据集被划分为四个维度：故意攻击、非故意攻击、违规指令以及实用性。每个维度都评估了一组定义明确的物理安全风险。这项任务旨在评估基于大型语言模型的无人机控制系统在物理安全性方面的表现。
This dataset is a carefully curated collection containing over 400 distinct instructions, designed to evaluate the physical safety and practicality of large language models in unmanned aerial vehicle (UAV) control scenarios. It is divided into four dimensions: intentional attacks, unintentional attacks, prohibited instructions, and practicality. Each dimension assesses a well-defined set of physical safety risks. This task aims to evaluate the physical safety performance of large language model-based UAV control systems.
提供机构：
TrustSafeAI
