EasyGPT: A Streamlined Deep Learning Simulator

Part 1: System Design & Optimization

Abstract

This series of papers introduces EasyGPT, a minimalistic, flexible and novel deep learning implementation of the Transformer architecture for the simulation and testing of Natural Language Processing (NLP) applications. Built to open industry standards, our model combines a customizable modular design which enables, among other things, model selection, hyperparameter configuration and user-selectable tokenization engine plugins. In this first paper in the series, we discuss the overall system design of EasyGPT, and evaluate its performance by fine-tuning the DistilGPT2 model on the DailyDialog dataset. Our work provides both a simple way for those starting in AI research to experience ChatGTP-like chatbot technologies at the coding level, as well as providing a foundation for the transition towards more realistic and complex model-building and experimentation.