[ethbr1]

Are there any good references for work on retraining large models to distinguish between control / system prompt and user data / prompt? (e.g. based on out-of-band type tagging of the former)