Anytime I see a prompt from one of these companies, I assume it matches the style of instructions the model encountered during pre-training.

And the kinds of instruction formats that were encountered during pre-training end up informing what style of instruction the model is best at following.

An extreme example would be prompt templates that a "raw" instruct-tuned LLM follows: the model will technically work with a suboptimal format, but you get much better performance if you follow the prompt template the model was trained/fine-tuned on.

-

It's not a guarantee a given prompting style was involved during pre-training of course, but at the very least it's going to a provide a jumping off point that the creators of the model co-signed on.