The startup's Invent a Dataset tool skips seed corpora and labeling guides to generate training sets.