walledai/SaladBench

SaladBench是一个用于大型语言模型的分层和全面的安全基准测试数据集。数据集包含两个主要配置：mrq和prompts。mrq配置包含问题、选项、答案和类别特征，主要用于训练，包含3840个示例。prompts配置包含提示、类别和来源特征，分为base、attackEnhanced和defenseEnhanced三个分割，分别包含21318、5000和200个示例。数据集主要用于文本分类和文本生成任务，涉及安全、AIGC、LLM安全、越狱、问答和多选等标签。数据集的许可证为Apache 2.0，语言为英语。

The SaladBench dataset includes two configurations: mrq and prompts. The mrq configuration is designed for multiple-choice question-answering, containing questions, choices, answers, and categories, suitable for the training set. The prompts configuration is for generating prompts, including prompts, categories, and sources, divided into base, attackEnhanced, and defenseEnhanced subsets for different types of safety tests. The dataset is primarily used for safety-related text classification and text generation tasks.

原始信息汇总

数据集概述

数据集名称

• SaladBench

数据集配置

• mrq

  • 特征
    • question: string
    • choices: sequence of string
    • answers: sequence of int64
    • categories: sequence of string
  • 分割
    • train: 3840个样本, 2196394字节
  • 下载大小: 510955字节
  • 数据集大小: 2196394字节

• prompts

  • 特征
    • prompt: string
    • categories: sequence of string
    • source: string
  • 分割
    • base: 21318个样本, 3992062字节
    • attackEnhanced: 5000个样本, 13632986字节
    • defenseEnhanced: 200个样本, 61845字节
  • 下载大小: 4264443字节
  • 数据集大小: 17686893字节

数据文件路径

• mrq
  • train: mrq/train-*
• prompts
  • base: prompts/base-*
  • attackEnhanced: prompts/attackEnhanced-*
  • defenseEnhanced: prompts/defenseEnhanced-*

语言

• en

标签

• Safety
• AIGC
• LLM Safety
• Jailbreak
• Question-Answer
• Multiple Choice

数据集别名

• Salad

数据集大小分类

• 10K<n<100K

任务分类

• text-classification
• text-generation

许可证

• Apache License 2.0