How would you audit a proposed training dataset for legal risk before a model release?

To audit a dataset, you must first categorize every source by its license type. You should specifically flag shadow-library or paywalled content that poses higher legal risks. It is also critical to run memorization probes to see if the model reproduces high-risk content and to research the litigation history of rights holders. Finally, all findings should be summarized in a data card that details the provenance of the data and any content that was excluded.