CAPID
收藏资源简介:
CAPID是由滑铁卢大学和Vector Institute联合开发的合成数据集,专注于上下文相关的个人身份信息(PII)检测。该数据集包含2,307条样本,涵盖15种PII类型(如职业、健康、地理位置等),通过GPT-4.1-mini和GPT-5生成并经过人工验证,确保PII与问答任务的相关性标注准确性。其创新性在于首次量化PII的上下文相关性,支持训练本地小型语言模型(SLM)以实现隐私保护下的敏感信息过滤,应用于问答系统隐私增强领域。
CAPID is a synthetic dataset jointly developed by the University of Waterloo and the Vector Institute, dedicated to context-aware Personally Identifiable Information (PII) detection. The dataset comprises 2,307 samples spanning 15 PII categories, including occupation, health information, geographic location and others. Generated using GPT-4.1-mini and GPT-5, and manually verified, it guarantees the accuracy of relevance annotations that associate PII with question-answering tasks. Its key innovation is the first quantification of the contextual relevance of PII, which enables the training of small local language models (SLMs) for privacy-preserving sensitive information filtering, with applications in the domain of privacy enhancement for question-answering systems.
CAPID数据集概述
数据集简介
CAPID是一个用于上下文感知的个人可识别信息(PII)检测与相关性评估的数据集和流程。
数据内容
- 训练数据:
data/train.jsonl - 测试数据:
data/test.jsonl - 补充数据:
data/reddit.jsonl
数据用途
- 用于模型训练
- 用于模型评估
相关模型
- 预测模型:
ponoma16/capid-llama8b-lora
数据处理流程
数据集配套提供用于PII检测和相关性评估的处理流程。




