Fine-tuning an LLM to Generate Lore Coherent Encounters for Dungeons and Dragons

A. Sivaganeshan, N. de Silva, and A. Peiris Proceedings of the 39th Pacific Asia Conference on Language, Information and Computation, 2025, pp. 710--722, Swathes of tasks that were erstwhile handled by other deep learning models are being taken over by Large Language Models (LLMs). While they may demonstrate reasonable results in the zero-shot configuration for most domains, in the contexts of more niche or esoteric domains, instruction tuning them on the domain at hand has shown to be effective and sometimes necessary. Dungeons & Dragons (D\&D) is the most commercially successful Tabletop Role Playing Game (TTRPG) with its own unique lore mostly set in the fantasy domain. The players are confronted with fantastical monsters in mathematically balanced encounters overcoming which, contribute to the calculation of player progress. Even with the plethora of information available for D\&D (or perhaps rather due to its abundance), the generation of an encounter that is coherent with the lore is a time-consuming and difficult process as there are no support tools available for the selection of coherent monsters. Recognizing this gap, we instruction-tuned a Mistral-based LLM that can function as an assistant on this matter, using instructions generated from publicly available D\&D datasets. Next, we conducted a number of prompt engineering experiments on the trained LLM, such that the output from the LLM would be a list of coherent monsters when a candidate monster is given in the input. The generated outputs were examined for the coherency of lore, theme, and environment. It was observed that the outputs were partially or fully coherent with the lore in about 66.0\% of the 241 candidate monsters tested.