I assume they have more/better task demonstrations for the multi-task finetuning phase. But that kind of data would be very easy to generate by calling their APIs. It's also possible to use a LLM to generate this kind of data from scratch, and even to do without RLHF by using Constitutional AI.