Review Book - Build a large language model

Written by: Tushar Sharma

Featured image for Review Book - Build a large language model

Below are my rough notes while reading the build-a-large-language-model-book.

LLM are deep neural networks trained on massive amount of text data. The "large" in LLM refers to both the models's size in terms of parameters and immense dataset on which it's trained.

LLM utilize an architecture called the transformer, which allows them to pay selective attention to different parts of the input when making predictions. Since LLM are capable pof generating text, LLM are also called generative AI. Transfer architecture was first introducted in 2017 paper called Attention is All You Need.

There are other variation of transformer architecture like BERT.

General process of creating LLM includes pretraining and fine-tuning. "pre" in pretraining refers to inital phase where a model like a LLM is trained on a large, diverse dataset (raw text). This pretrained model then serves as a foundation resource that can be further refined through fine-tuning, a process where model is specifically trained on a narrow dataset for a specific task.

Fine-tuning has two categories

instruction fine tuning: labelled dataset consist of instruction and answer pairs.
classification fine tuning: labelled dataset consist of texts and associated class label.

GPT stands for generative pretrained transformer. GPT models are adept at :

Zero shot learning: Generalize a completely unseen task
Few shot learning: Learning from minimal number of examples which the user provides