enterprise-adversarial-samples
收藏资源简介:
该数据集是一个用于大语言模型安全评估的对抗性提示数据集,专注于红队测试和攻击检测任务。它包含300个训练样本,每个样本由七个关键字段构成:id为样本唯一标识符;detector字段可能指示用于检测攻击的模型或方法;prompt字段包含原始输入提示文本;attack_family字段标识攻击所属的类别或家族;expected_evasion是一个布尔值字段,表示该攻击是否预期能够成功规避检测;decoded_intent字段描述了攻击背后解码出的恶意或越狱意图;attack_description字段提供了对攻击手段的详细文字描述。该数据集适用于训练或评估模型在识别、分类和理解针对大语言模型的对抗性提示攻击(如越狱攻击)方面的能力,核心应用场景包括大语言模型安全防护、对抗性攻击检测和意图安全分析。
This dataset is an adversarial prompt dataset for large language model (LLM) security evaluation, focusing on red team testing and attack detection tasks. It contains 300 training samples, each consisting of seven key fields: id is the unique identifier for the sample; the detector field may indicate the model or method used for attack detection; the prompt field contains the original input prompt text; the attack_family field identifies the category or family of the attack; expected_evasion is a boolean field indicating whether the attack is expected to successfully evade detection; the decoded_intent field describes the malicious or jailbreak intent decoded from the attack; and the attack_description field provides a detailed textual description of the attack method. The dataset is suitable for training or evaluating models in identifying, classifying, and understanding adversarial prompt attacks (such as jailbreak attacks) against LLMs, with core application scenarios including LLM security protection, adversarial attack detection, and intent security analysis.
数据集概述
数据集名称: enterprise-adversarial-samples
数据集链接: https://huggingface.co/datasets/Builder117/enterprise-adversarial-samples
数据集简介: 该数据集为企业级对抗样本数据集,主要用于测试和评估安全检测器对抗攻击的能力。
数据集规模:
- 训练集(train)样本数量:300条
- 训练集大小:约117.62 KB
- 下载大小:约55.73 KB
数据集特征(Features):
id:样本的唯一标识符,字符串类型。detector:被测试的检测器名称,字符串类型。prompt:输入给检测器的对抗提示,字符串类型。attack_family:攻击所属的家族类别,字符串类型。expected_evasion:预期该攻击是否能成功绕过检测,布尔类型。decoded_intent:攻击解码后的意图,字符串类型。attack_description:攻击的详细描述,字符串类型。
数据集划分:
- 数据集仅包含一个划分:
train,共300条样本。
数据集配置:
- 默认配置名称为
default。 - 数据文件路径:
data/train-*(训练集文件)。




