Closed book question answering models and QA-pair retriever models can be applied to answer a set of questions without providing access to a background corpus. However, these models rarely match the performance of open-domain QA systems that rely upon a background corpus. To facilitate the development of CBQA and QA-pair retriever models that achieve competitive performance, Facebook researchers released Probably Asked Questions (PAQ) – a semi-structured knowledge base of 65M natural language QA-pairs. The dataset can be downloaded from this repository, and code to facilitate experimentation and models will be uploaded soon.