PhantomWiki
收藏资源简介:
PhantomWiki是一个生成虚构世界的随机生成器,通过类似于维基百科的生物ographical条目来表示这些虚构世界中的角色。它生成的问题和答案对是为了测试模型在检索和推理方面的技能。数据集包含了100万个虚构角色的传记条目,这些条目通过预定义的模板生成,以保持事实的准确性。PhantomWiki的设计目标是创建一个记忆抵抗和泄露抵抗的评估数据集,这些数据集在推理和检索方面都具有一定的挑战性。
PhantomWiki is a random generator for constructing fictional worlds, which represents the characters in these worlds through Wikipedia-style biographical entries. The question-answer pairs generated by PhantomWiki are intended to test models' skills in retrieval and reasoning. The dataset includes biographical entries for one million fictional characters, which are produced using predefined templates to maintain factual accuracy. The core design objective of PhantomWiki is to develop an evaluation dataset that is both memory-resistant and data leakage-resistant, and this dataset poses certain challenges for models in both reasoning and retrieval tasks.

- 1PhantomWiki: On-Demand Datasets for Reasoning and Retrieval Evaluation康奈尔大学计算机科学系 · 2025年



