when training a LLM would you prefer to first expose it to traumatizing content and then teach it to suppress the knowledge in its responses, or would you keep it innocent by removing the dark damaging topics from the training data?