Ah, that makes more sense, thank you. Since this was mentioned in the tokenizer section and the number of unique tokens wasn't mentioned I misunderstood.