/para_normalized_losses/training