The Small Language Model from Microsoft, called Phi-3, was trained using a novel dataset called TinyStories.