Hands-On Large Language Models

Chapter 11. Fine-Tuning Representation Models for Classification

In Chapter 4, we used pretrained models to classify our text. We kept the pretrained models as they were without any modifications to them. This might make you wonder, what happens if we were to fine-tune them?

If we have sufficient data, fine-tuning tends to lead to some of the best-performing models possible. In this chapter, we will go through several methods and applications for fine-tuning BERT models. “Supervised Classification” demonstrates the general process of fine-tuning a classification model. Then, in “Few-Shot Classification”, we look at SetFit, which is a method for efficiently fine-tuning a high-performing model using a small number of training examples. In “Continued Pretraining with Masked Language Modeling”, we will explore how to continue training a pretrained model. Lastly, classification on a token level is explored in “Named-Entity Recognition”.

We will focus on nongenerative tasks, as generative models will be covered in Chapter 12.

Supervised Classification

In Chapter 4, we explored supervised classification tasks by leveraging pretrained representation models that were either trained to predict sentiment (task-specific model) or to generate embeddings (embedding model), as shown in Figure 11-1.

Figure 11-1. In Chapter 4, we used pretrained models to perform classification without updating ...