Comment by simonw

I don't see why they would do that. The whole point of training a model like Gemini is that you train the model - if they want it to work great against those different categories of document the likely way to do it is to add a whole bunch of those documents to Gemini's regular training set.