Yes, this can work. I’ve done that in a few cases.

In fact, if you split data preprocessing in small enough steps, they could also be run on weaker LLMs. It would take a lot more time, but that is doable.