At the moment, this is relying on the dataset cards, so the similarity does indeed work better for longer dataset cards. I plan for a version that will directly use the dataset to create the similarity scores, which should hopefully work better!