Maybe set up the training code so different foundation models can be plugged in for finetuning and the it’s just compute if somebody wants a different starting model.

Note there are free interfaces to these models such as https://spellbook.scale.com/ . Also note there is a lot of data collected out there already.