Reducing dropped tokens could also improve model training by reducing gradient noise