They are trained on some mix with minimal fraction of math. That's how it was from the beginning. But we can rebalance it by adding quality generated content. Just content will cost millions of $$ to generate. Distillation on new level looks like logical next step.