I don't think it's necessarily about the parameter count, but the amount of training material about the Bible relative to the rest of the training material, wit

I don't think it's necessarily about the parameter count, but the amount of training material about the Bible relative to the rest of the training material, with higher parameter models able to retain more Bible information with a higher proportion of training on other topics.