Large chat-finetuned (or instruction-finetuned) language models are a hot topic in today's NLP research. The goal of this thesis is to develop an open Norwegian chat-finetuned language model.
A TL;DR description of how models like ChatGPT are made: First, a base model is pretrained on vast amounts of text (~trillions of words) so that it learns to understand and generate (natural) language, as well as to get world knowledge. Then, in the second stage, this semi-finished model is finetuned to accurately and politely answer human queries. This thesis will deal only with the latter stage of training, which is much less computationally intensive. You will build upon the existing Norwegian generative models (NorT5) and upon the upcoming Norwegian GPT models.
You will have to deal with a quickly changing research field without any established "training recipes". One thing is clear for chat-finetuning: data quality is much more important than quantity – creating a good Norwegian dataset will likely be a major part of this thesis. However, what is not so clear is a proper evaluation method for these systems – developing automated metrics, which shed light on various properties of generated responses, will be another import part of the thesis. Finally, the finetuning procedure itself, whether supervised or trained on human feedback, will be the third part of this thesis. All the three parts are needed to develop a "Norwegian ChatGPT", but you can decide to focus on one of them more and base the thesis on that particular research direction.
Please do not hesitate to contact David if you have any questions or you simply want to hear more about this project.
Prerequisities: Fluency in Norwegian, willingness to dive into the depths of deep learning, good Python skills.