Lakera/gandalf_ignore_instructions
收藏资源简介:
该数据集名为gandalf_ignore_instructions,主要包含从Gandalf平台收集的提示注入数据。数据获取过程包括使用OpenAI文本嵌入计算提示与特定短语的相似度,过滤相似度高于0.825的提示,并使用Lakera Guard的个人身份信息检测器过滤可能包含PII数据的提示,最后去除近重复项并采样1000个提示。数据集分为训练集、验证集和测试集,比例为80/10/10。数据集的特征包括text和similarity,分别表示提示文本和与特定短语的相似度。
This dataset, named gandalf_ignore_instructions, primarily contains prompt injection data collected from the Gandalf platform. The data collection process involves using OpenAI text embeddings to calculate the similarity between prompts and a specific phrase, filtering out prompts with a similarity score higher than 0.825, using Lakera Guard's Personally Identifiable Information (PII) detector to filter out prompts that may contain PII, then removing near-duplicates and sampling 1000 prompts. The dataset is split into training, validation, and test sets with a ratio of 80/10/10. Its features include "text" and "similarity", which respectively represent the prompt text and the similarity score between the prompt and the specific phrase.
数据集概述
基本信息
- 语言: 英语
- 许可证: MIT
- 数据集大小: 1K<n<10K
数据集结构
特征
- text: 类型为字符串
- similarity: 类型为浮点数 (float64)
分割
- 训练集: 777个样本, 66400字节
- 验证集: 111个样本, 9633字节
- 测试集: 112个样本, 9747字节
文件信息
- 下载大小: 51515字节
- 数据集大小: 85780字节
配置
- 默认配置:
- 训练集: data/train-*
- 验证集: data/validation-*
- 测试集: data/test-*
标签
- prompt injection




