kilian-group/phantom-wiki-v1

Dataset Card for PhantomWiki

This repository contains pre-generated instances of the PhantomWiki dataset, created using the phantom-wiki Python package. PhantomWiki is a framework for evaluating LLMs, particularly RAG and agentic workflows, designed to be resistant to memorization. Unlike fixed datasets, PhantomWiki generates unique instances on demand, ensuring novelty and preventing data leakage.

Dataset Details

Dataset Description

PhantomWiki generates a synthetic fictional universe populated with characters and facts, mirroring the structure of a fan wiki. These facts are reflected in a large-scale corpus, and diverse question-answer pairs of varying difficulties are then generated.

• Created by: Albert Gong, Kamilė Stankevičiūtė, Chao Wan, Anmol Kabra, Raphael Thesmar, Johann Lee, Julius Klenke, Carla P. Gomes, Kilian Q. Weinberger
• Funded by: See the paper for details.
• License: MIT License
• Paper: PhantomWiki: On-Demand Datasets for Reasoning and Retrieval Evaluation
• Code: https://github.com/kilian-group/phantom-wiki

Dataset Sources

• Repository (Code): https://github.com/kilian-group/phantom-wiki

Uses

PhantomWiki is designed to evaluate retrieval augmented generation (RAG) systems and agentic workflows. To avoid data leakage and overfitting, generate a new, unique PhantomWiki instance for each evaluation. Our paper details an analysis of frontier LLMs using PhantomWiki.

Dataset Structure

PhantomWiki provides three configurations:

1. question-answer: Question-answer pairs generated using a context-free grammar
2. text-corpus: Documents generated using natural-language templates
3. database: Prolog database containing the facts and clauses representing the universe

Each universe is saved as a separate split. See the original repository for details on generation and usage.

Bias, Risks, and Limitations

PhantomWiki, while effective at evaluating complex reasoning and retrieval, is limited in its representation of family relations and attributes. Extending its complexity is a future research direction. For a holistic evaluation, combine PhantomWiki with other benchmarks.

Citation

@article{2025_phantomwiki,
  title={{PhantomWiki: On-Demand Datasets for Reasoning and Retrieval Evaluation}},
  author={Albert Gong and Kamilė Stankevičiūtė and Chao Wan and Anmol Kabra and Raphael Thesmar and Johann Lee and Julius Klenke and Carla P. Gomes and Kilian Q. Weinberger},
  year={2025},
  journal={todo},
  url={todo},
  note={Under Review},
}