Hi! I work in NLP. I'm curious about finding use cases for GPT-3 like models, but an issue I have is that these models sometimes produce garbage, which is dangerous.

How do you do it? Did you manage to somehow plug in a verification model for its output? What were the results?