Type what you care about ("does the model know drug recall dates", "quotes people get wrong", "startup ownership facts") and this finds the datasets and the actual verified claims closest in meaning, using an embedding model, not word matching.