Series of 125m parameter text generation models trained from scratch