In this article, I will show how to easily train GPT-class neural networks from home. Let me start by saying that we won’t train NN from scratch, as that would require 8 (eight!) A100-class GPUs at least and a massive dataset. Instead, we’ll focus on fine-tuning a pre-trained GPT-2 model using a smaller dataset, which anyone can easily make or find online. OpenAI has kindly released GPT-2 under Modified MIT License.
We’ll use the nanoGPT repository created by Andrej Karpathy for fast and easy GPT training. He has a comprehensive video lecture explaining how GPT-2 works and how to train such a neural network. However, we’re interested in fine-tuning the model using our own dataset and seeing the difference from the original (GPT-2 trained by OpenAI).