HalluRAG

HalluRAG is a dataset designed to train classifiers on closed-domain hallucinations for queries and knowledge that refers to information which has not been available during training of the LLM. This is realized through a focus on recency which was employed to ensure the information emerged after a cut-off date. The texts are from timestamped Wikipedia articles which we used to verify that information is recent and was not available—even in an older form or relying on similar older content—during training. Wikipedia provides a semi-structured source of information with a wide variety of topics and a large number of references within its articles.