This work has been created as part of the Seminar Machine Learning at the Technical University of Munich. Throughout this seminar two to three students had to present a paper about foundational/generative AI models each week, which was then followed by a discussion in the plenum. My paper was Stanford Alpaca. Unhappy with the original evaluation of the Alpaca paper, I decided to not only talk about the paper and its success but focussed the majority of my work on a critical evaluation of Alpaca. This repository contains the slides and the paper I wrote for the seminar.
This paper introduces Alpaca, a novel instruction-tuned language model that aims to challenge established models like ChatGPT. Unlike existing models, Alpaca is openly available, compact in size, and can be trained at a significantly lower cost of just 600$. Despite its cost-effectiveness, Alpaca strives to deliver comparable performance to ChatGPT. The paper delves into the internal workings of Alpaca and conducts a comprehensive evaluation to determine whether it can match the claims of performance comparable to ChatGPT.