The CBT dataset is constructed from text paragraphs and corresponding questions, and the question-answering data are all from books provided free of charge by the Gutenberg Project. This dataset is used to directly measure language models and a wider language environment for question-answering and simulation search.

The CBT dataset was released by Facebook in 2016. The main publishers are Felix Hill, Antoine Bordes, Sumit Chopra and Jason Weston. The related paper is "The Goldilocks Principle: Reading Children's Books with Explicit Memory Representations".