A parks desk extracts trail-hazard mentions from ranger notes. Staff can freeze the pretrained transformer and train a light head, or they can update more of the stack. What should that experiment measure?
Select an answer to reveal the explanation.
Short Explanation
Freeze the transformer and train a light head, or update more of the stack. Compare those two on validation entity quality and memory cost. How serious the extra training looks, GPU hours, and fluent ranger notes are not the metric.
Full Explanation
Frozen versus updated transformer weights is a measured NLP-task factor. A light head on a frozen stack can be enough for entity mentions, and updating more of the stack can help or can waste memory. The comparison belongs on validation entity quality and the memory or time bill. How serious the training looks is not a metric.