Why is customizing FMs important?

In the previous chapter, we looked at several prompt engineering techniques to improve the performance of a model. As we also saw in Chapter 1 (and shown in Figure 4.1), these FMs are trained on massive amounts of data (GBs, TBs, or PBs) with millions to billions of parameters, allowing them to understand relationships between words in context to predict subsequent sequences:

Figure 4.1 – Training an FM

So, why do we need to customize these models?

That’s a fair question since a lot of use cases can be directly solved by using prompt engineering and RAG techniques (which we will cover in Chapter 5). However, consider a situation where you require the model to adhere to a particular writing style, output format, or domain-specific terminology. For instance, you may need the model to analyze financial earnings reports or medical records accurately. In such cases, the pre-trained models might not have been...