Tag
This paper demonstrates that scaling laws fit on small transformer models can accurately predict the loss of much larger models trained on particle physics jet data, enabling compute budgets to be translated into expected physics performance before large training runs. They release five pretrained models and the full training recipe.