I'd think it's really good at tokenization to word. That's exactly what it's trained to do.