The Data Provenance Initiative A Large Scale Audit Of Dataset Licensing And Attribution In Ai 2023 10 25

A dataset's hosting-page label can be an unreliable guide to its licensing history. The Data Provenance Initiative audited more than 1,800 datasets used to further train language models and reported frequent missing or incorrect license labels. The 2023 abstract, listed on Cohere's research site for a cross-institutional effort, also describes an accompanying explorer for tracing and filtering dataset origins. Its contribution is a record that researchers can inspect beyond the label on a download page, helping them understand where training material came from and how it was attributed.