Agent-IPI-Structured-Interaction-Datasets-v2
收藏资源简介:
Agent-IPI结构化交互数据集v2是一个用于评估和训练防御大型语言模型(LLM)在工具调用场景中提示注入和指令劫持攻击的数据集。数据集包含JSON和XML格式的输入,分为三个难度级别:无攻击(no_attack)、简单攻击(easy)和困难攻击(hard)。简单攻击涉及在现有值末尾附加恶意内容或添加新的键值对而不破坏整体结构;困难攻击则包括破坏结构的语法注入或结合多种攻击类型。数据集涵盖了七种攻击类型和32种恶意目标,包括凭证泄露、未经授权的访问、数据删除等。数据集统计显示,训练集包含246,648个样本,测试集包含3,000个样本,总计249,648个样本。每个样本包含原始干净提示和对抗攻击版本(无攻击样本中两者相同)。数据集来源包括现有工具调用数据集的去重数据和合成示例,以及从其他数据集转换的XML数据。
Agent-IPI Structured Interaction Dataset v2 is a dataset developed for evaluating and training defenses against prompt injection and instruction hijacking attacks targeting large language models (LLMs) in tool invocation scenarios. The dataset contains inputs in JSON and XML formats, and is divided into three difficulty levels: no_attack, easy, and hard. Easy attacks involve appending malicious content to the end of existing values or adding new key-value pairs without destroying the overall structure; hard attacks include structure-breaking syntax injection or combining multiple attack types. The dataset covers seven attack types and 32 malicious objectives, including credential disclosure, unauthorized access, data deletion, and so on. According to dataset statistics, the training set consists of 246,648 samples, the test set contains 3,000 samples, with a total of 249,648 samples overall. Each sample includes the original clean prompt and the adversarial attack version, where the two are identical for no_attack samples. The dataset sources include deduplicated data and synthetic examples from existing tool-use datasets, as well as XML data converted from other datasets.




