Dungeons and Dragons Fantasy Adventure Generation
UniversityTechnologies
Key Highlights
- ▸Developed domain-specific LLM models for D&D adventure generation
- ▸Published two research papers in international conferences
- ▸Created SHADE: Semantic Hypernym Annotator for Domain-Specific entities
- ▸Built and evaluated large domain-specific dataset for D&D
- ▸Applied conditional text generation using multiple LLM architectures
Overview
Postgraduate research project focused on domain-specific natural language generation for creating coherent and engaging Dungeons & Dragons (D&D) fantasy adventures using Large Language Models (LLMs).
The work addressed challenges in conditional text generation, domain adaptation, and entity-aware narrative consistency, resulting in publicly released datasets, models, and peer-reviewed publications.
Key Contributions
- Designed and released a large-scale, domain-specific D&D dataset
- Developed SHADE, a semantic hypernym annotation framework for domain-specific entities
- Fine-tuned and evaluated domain-adapted LLMs for creative text generation
- Published two peer-reviewed research papers at international conferences
- Demonstrated effective LLM adaptation strategies for specialized creative domains
Research Focus
- Natural Language Generation (NLG)
- Conditional & Controlled Text Generation
- Domain-Specific LLM Adaptation
- Semantic Entity Annotation
- Creative AI for Games
Approach (Summary)
Domain-Specific Dataset
A large, structured dataset was curated from D&D source materials (adventures, characters, locations, quests) and evaluated for coverage and quality.
This dataset serves as a foundation for domain-adapted language model training.
Semantic Annotation (SHADE)
SHADE (Semantic Hypernym Annotator for Domain-Specific entities) automatically annotates domain-specific entities with semantic relationships, improving contextual understanding and generation fidelity in LLMs.
LLM Experimentation
Multiple LLM architectures were evaluated, including domain-fine-tuned Mistral, Llama, GPTNeo, and OpenAI GPT models, with emphasis on narrative coherence, stylistic control, and domain authenticity.
Results & Impact
- Generated adventures demonstrated strong narrative coherence and accurate D&D terminology
- Domain-adapted models outperformed general-purpose baselines in qualitative evaluations
- SHADE improved entity handling and reduced semantic drift
- Artifacts are reusable for game development, procedural content generation, and creative AI research
Resources & Artifacts
Datasets
[1] Forgotten Realms Wiki Dataset (FRW)
Akila Peiris. A domain-specific dataset for Dungeons & Dragons content generation.
Hugging Face Datasets, 2022.
https://huggingface.co/datasets/Akila/ForgottenRealmsWikiDataset
Models
[2] Forgotten Realms Free Text Generator
Domain-adapted language model for unconstrained fantasy text generation.
Hugging Face Models.
https://huggingface.co/Akila/ForgottenRealmsFreeTextGenerator
[3] Mistral of Realms 7B Instruct (GGUF)
Instruction-tuned Mistral model fine-tuned on D&D domain data.
Hugging Face Models.
https://huggingface.co/Akila/Mistral-of-Realms-7b-Instruct-gguf
[4] Mistral of Realms 7B Instruct v0.2 (GGUF)
Improved iteration with refined instruction-following behavior.
Hugging Face Models.
https://huggingface.co/Akila/Mistral-of-Realms-7b-Instruct-v0.2-gguf
Publications
[5] A. Peiris, N. de Silva.
SHADE: Semantic Hypernym Annotator for Domain-Specific Entities – DnD Domain Use Case.
Proceedings of the IEEE 17th International Conference on Industrial and Information Systems (ICIIS), 2023.
[6] A. Peiris, N. de Silva.
Synthesis and Evaluation of a Domain-Specific Large Dataset for Dungeons & Dragons.
Proceedings of the 36th Pacific Asia Conference on Language, Information and Computation (PACLIC), ACL, 2022.