A research team is tasked with building a system to classify legal documents into highly specific categories. They have a large, well-labeled dataset. The team is debating between two primary approaches: (1) fine-tuning a general-purpose large language model, and (2) training a classification head on top of a pre-trained encoder model. Compare and contrast these two approaches in the context of this specific task. Your evaluation should consider potential differences in performance, computational requirements, and ease of implementation.