A development team starts with a large, pre-trained language model that has a general understanding of text. Their goal is to create a system that can classify customer feedback emails into one of three categories: 'Urgent', 'Standard', or 'Spam'. To adapt the general-purpose model for this specific classification task, what is the most appropriate and standard architectural change they should implement?

Adapting a Pre-trained Model for Different NLP Tasks

A research team has access to a powerful, pre-trained language model that produces a contextualized numerical representation for every token in an input sequence. To solve specific problems, they must add a new, task-specific prediction network (a 'head') on top of this pre-trained base. Match each downstream task with the architectural design of the prediction head best suited to accomplish it.