SAFETEXT

SAFETEXT是由加利福尼亚大学圣巴巴拉分校和哥伦比亚大学合作创建的第一个常识物理安全数据集，包含367个真实生活场景，每个场景配有一对安全和物理不安全的建议。数据集通过从Reddit收集人类编写的帖子，并经过五个阶段的筛选和重写过程创建。SAFETEXT旨在解决自然语言处理中模型生成不安全文本的问题，特别是在需要常识推理以理解可能导致物理伤害的文本方面。该数据集适用于文本生成和常识推理任务，有助于评估和提高机器学习模型在处理日常生活中的安全问题时的性能。SAFETEXT is the first commonsense physical safety dataset co-developed by the University of California, Santa Barbara and Columbia University. It contains 367 real-life scenarios, each paired with a pair of safe and physically unsafe recommendations. The dataset was constructed by collecting human-authored posts from Reddit and undergoing a five-stage filtering and rewriting process. SAFETEXT aims to address the issue of unsafe text generation by natural language processing models, particularly in scenarios requiring commonsense reasoning to understand text that may lead to physical harm. This dataset is applicable to text generation and commonsense reasoning tasks, and helps evaluate and improve the performance of machine learning models when handling safety-related issues in daily life.

搜集汇总

背景与挑战

背景概述

SAFETEXT是首个常识物理安全数据集，由加利福尼亚大学圣巴巴拉分校和哥伦比亚大学合作创建，包含367个真实生活场景，每个场景配有一对安全和物理不安全的建议。该数据集旨在解决自然语言处理中模型生成不安全文本的问题，特别关注需要常识推理以避免物理伤害的文本，适用于评估和提高模型在文本生成和常识推理任务中的安全性能。