FIREBALL

FIREBALL是一个包含近25,000个独特游戏会话的大型数据集，源自Discord上的真实Dungeons & Dragons游戏，并包含真实的游戏状态信息。该数据集记录了使用Avrae bot进行游戏的玩家会话，捕捉了语言、游戏命令和底层游戏状态信息。FIREBALL数据集能够通过使用Avrae状态信息改善自然语言生成（NLG），不仅提高了自动化指标，还提升了人类对质量的判断。此外，该数据集还展示了经过微调的大型语言模型（LLMs）能够生成可执行的Avrae命令。FIREBALL数据集的应用领域包括改善游戏体验、辅助新手玩家，以及通过提供结构化的游戏状态信息来推动更复杂的创意生成任务，如一致的长期叙事生成和基于游戏状态生成叙述和对话的视频游戏。

FIREBALL is a large-scale dataset containing nearly 25,000 unique game sessions, sourced from real Dungeons & Dragons games hosted on Discord with authentic game state information included. This dataset documents player sessions of games run via the Avrae bot, capturing natural language utterances, game commands, and underlying game state data. The FIREBALL dataset facilitates improved natural language generation (NLG) by leveraging Avrae-provided game state information: it not only elevates automated evaluation metrics but also improves human-rated quality assessments. Furthermore, this dataset validates that fine-tuned large language models (LLMs) can generate executable Avrae commands. Applications of the FIREBALL dataset span enhancing gaming experiences, assisting novice players, and enabling more sophisticated creative generation tasks by supplying structured game state information—including generating coherent long-form narratives, and producing video game narration and dialogue grounded in in-game states.