kilian-group/phantom-wiki-v050

PhantomWiki 是一个用于生成虚构宇宙和事实的大型语料库数据集，包含多种问题-答案对。该数据集是使用 `phantom-wiki` Python 包动态生成的，不是基于任何现有数据。它由三个主要组件组成：`question-answer`（包含可调难度的问题-答案对）、`text-corpus`（包含模仿粉丝网站风格的文档）和 `database`（包含表示宇宙的事实和条款的 Prolog 数据库）。PhantomWiki 旨在评估检索增强生成（RAG）系统和代理工作流程，其合成和可控的性质使其特别适合于解开大型语言模型中的推理和检索能力。

PhantomWiki is a dataset that generates a fictional universe and facts, containing diverse question-answer pairs. It is created dynamically using the `phantom-wiki` Python package and is not based on any existing data. The dataset consists of three main components: `question-answer` (with tunable difficulty question-answer pairs), `text-corpus` (documents mimicking the style of fan-wiki websites), and `database` (a Prolog database with facts and clauses representing the universe). PhantomWiki is designed to evaluate retrieval augmented generation (RAG) systems and agentic workflows, and its synthetic and controllable nature is particularly useful for disentangling reasoning and retrieval capabilities in large language models.