Dungeons and Dragons fantasy adventure generation

Abstract

In this thesis, we evaluate the possibility of creating an automated fantasy adventure generation system for the popular tabletop game Dungeons and Dragons. We have evaluated what the existing practices are when story generation is considered and their limitations, especially given the domain specificity. We have detailed our steps in creating the Forgotten Realms Wiki (FRW) data set, a Dungeons and Dragons domainspecific data set containing several sub-datasets with various degrees of pre-processing useful in different Natural Language Processing tasks. Going a step further, we have evaluated the goodness of these data via semantic similarity comparisons, establishing the uniqueness and validity of the existence of these different data sets. We have created multiple text generation models capable of free text generation as well as in-filling focused text generation. In addition, the study describes the potential pitfalls when using uni-directional models such as attention-based language models for infilling tasks. We have finally created an instruction fine-tuned Large Language Model, Mistral of Realms Instruct 7b using FRW data set. Then we have created a system comprised of Artificial Intelligence agents backed by this instruct-fine tuned model to create a system capable of generating the necessary components of a dungeons and dragons adventure module including the main narrative, character descriptions, and gameplay mechanics. We have qualitatively evaluated the goodness of the generated content with a combination of similar techniques.