About
Imagine if a GPT language model could learn how to debate like an Australian Federal politician in the House of Representatives? Now imagine if it could debate itself? Well, you'd get Always Question Time, providing around-the-clock, on-demand coverage of Question Time from the House of Representatives! Remember — all names, questions, answers and content portrayed, while seemlingly real and sometimes hillarious, are hallucinated in the mind of a GPT language model. No identification with actual persons, places, bills or legislation is intended or should be inferred.
Demo site
http://alwaysquestiontime.com/
How was this created?
The language model was trained using Andrej Karpathy's excellent llama2.c framework. The training data was sourced from Question Time transcripts (via tveeder.com), and the official Hansard from the House of Representatives. Many thousands of questions and answers from hundreds of Parliamentary seatings, spanning the period of 2018 to 2023, were fed into the GPT to create a completion model.
Why?
My main objective was to explore the world of language models, and to have some fun with the large corpus of tveeder.com data.