There's no dataset to "steer to" or "match". LLMs don't memorize the vast majority of what they train on. In the simplest form, the LLM has learnt that these k

There's no dataset to "steer to" or "match". LLMs don't memorize the vast majority of what they train on. In the simplest form, the LLM has learnt that these kind of "emotional inputs" change the output in a meaningful way. It has learnt how to model this change. For all intents and purposes, it is an emotional trigger.