What metrics would you monitor for an NLP model deployed in production at Berkeley Research Group, and how would you detect drift?
