So maybe the training data has a lot of old English writing and overcoming the model's tendency to use em dashes everywhere with custom instructions would use up more electricity.