---
license: mit
task_categories:
- question-answering
language:
- en
size_categories:
- 1M<n<10M
configs:
- config_name: database
  data_files:
  - split: depth_20_size_50_seed_1
    path: database/depth_20_size_50_seed_1-*
  - split: depth_20_size_50_seed_2
    path: database/depth_20_size_50_seed_2-*
  - split: depth_20_size_50_seed_3
    path: database/depth_20_size_50_seed_3-*
  - split: depth_20_size_500_seed_1
    path: database/depth_20_size_500_seed_1-*
  - split: depth_20_size_500_seed_2
    path: database/depth_20_size_500_seed_2-*
  - split: depth_20_size_500_seed_3
    path: database/depth_20_size_500_seed_3-*
- config_name: question-answer
  data_files:
  - split: depth_20_size_50_seed_1
    path: question-answer/depth_20_size_50_seed_1-*
  - split: depth_20_size_50_seed_2
    path: question-answer/depth_20_size_50_seed_2-*
  - split: depth_20_size_50_seed_3
    path: question-answer/depth_20_size_50_seed_3-*
  - split: depth_20_size_500_seed_1
    path: question-answer/depth_20_size_500_seed_1-*
  - split: depth_20_size_500_seed_2
    path: question-answer/depth_20_size_500_seed_2-*
  - split: depth_20_size_500_seed_3
    path: question-answer/depth_20_size_500_seed_3-*
  - split: depth_20_size_5000_seed_1
    path: question-answer/depth_20_size_5000_seed_1-*
- config_name: text-corpus
  data_files:
  - split: depth_20_size_50_seed_1
    path: text-corpus/depth_20_size_50_seed_1-*
  - split: depth_20_size_50_seed_2
    path: text-corpus/depth_20_size_50_seed_2-*
  - split: depth_20_size_50_seed_3
    path: text-corpus/depth_20_size_50_seed_3-*
  - split: depth_20_size_500_seed_1
    path: text-corpus/depth_20_size_500_seed_1-*
  - split: depth_20_size_500_seed_2
    path: text-corpus/depth_20_size_500_seed_2-*
  - split: depth_20_size_500_seed_3
    path: text-corpus/depth_20_size_500_seed_3-*
dataset_info:
- config_name: database
  features:
  - name: content
    dtype: string
  splits:
  - name: depth_20_size_50_seed_1
    num_bytes: 25163
    num_examples: 1
  - name: depth_20_size_50_seed_2
    num_bytes: 25205
    num_examples: 1
  - name: depth_20_size_50_seed_3
    num_bytes: 25015
    num_examples: 1
  - name: depth_20_size_500_seed_1
    num_bytes: 191003
    num_examples: 1
  - name: depth_20_size_500_seed_2
    num_bytes: 190407
    num_examples: 1
  - name: depth_20_size_500_seed_3
    num_bytes: 189702
    num_examples: 1
  download_size: 192917
  dataset_size: 646495
- config_name: question-answer
  features:
  - name: id
    dtype: string
  - name: question
    dtype: string
  - name: intermediate_answers
    dtype: string
  - name: answer
    sequence: string
  - name: prolog
    struct:
    - name: query
      sequence: string
    - name: answer
      dtype: string
  - name: template
    sequence: string
  - name: type
    dtype: int64
  - name: difficulty
    dtype: int64
  splits:
  - name: depth_20_size_50_seed_1
    num_bytes: 299559
    num_examples: 500
  - name: depth_20_size_50_seed_2
    num_bytes: 303664
    num_examples: 500
  - name: depth_20_size_50_seed_3
    num_bytes: 293959
    num_examples: 500
  - name: depth_20_size_500_seed_1
    num_bytes: 308562
    num_examples: 500
  - name: depth_20_size_500_seed_2
    num_bytes: 322956
    num_examples: 500
  - name: depth_20_size_500_seed_3
    num_bytes: 300467
    num_examples: 500
  - name: depth_20_size_5000_seed_1
    num_bytes: 338703
    num_examples: 500
  download_size: 453442
  dataset_size: 2167870
- config_name: text-corpus
  features:
  - name: title
    dtype: string
  - name: article
    dtype: string
  - name: facts
    sequence: string
  splits:
  - name: depth_20_size_50_seed_1
    num_bytes: 25754
    num_examples: 51
  - name: depth_20_size_50_seed_2
    num_bytes: 26117
    num_examples: 50
  - name: depth_20_size_50_seed_3
    num_bytes: 25637
    num_examples: 51
  - name: depth_20_size_500_seed_1
    num_bytes: 262029
    num_examples: 503
  - name: depth_20_size_500_seed_2
    num_bytes: 260305
    num_examples: 503
  - name: depth_20_size_500_seed_3
    num_bytes: 259662
    num_examples: 504
  download_size: 275933
  dataset_size: 859504
---

# Dataset Card for PhantomWiki

**This repository is a collection of PhantomWiki instances generated using the `phantom-wiki` Python package.**

PhantomWiki is framework for evaluating LLMs, specifically RAG and agentic workflows, that is resistant to memorization.
Unlike prior work, it is neither a fixed dataset, nor is it based on any existing data.
Instead, PhantomWiki generates unique dataset instances, comprised of factually consistent document corpora with diverse question-answer pairs, on demand.

## Dataset Details

### Dataset Description

PhantomWiki generates a fictional universe of characters along with a set of facts.
We reflect these facts in a large-scale corpus, mimicking the style of fan-wiki websites.
Then we generate question-answer pairs with tunable difficulties, encapsulating the types of multi-hop questions commonly considered in the question-answering (QA) literature.

- **Created by:** Albert Gong, Kamilė Stankevičiūtė, Chao Wan, Anmol Kabra, Raphael Thesmar, Johann Lee, Julius Klenke, Carla P. Gomes, Kilian Q. Weinberger
- **Funded by:** AG is funded by the NewYork-Presbyterian Hospital; KS is funded by AstraZeneca; CW is funded by NSF OAC-2118310; AK is partially funded by the National Science Foundation (NSF), the National Institute of Food and Agriculture (USDA/NIFA), the Air Force Office of Scientific Research (AFOSR), and a Schmidt AI2050 Senior Fellowship, a Schmidt Sciences program.
- **Shared by \[optional\]:** \[More Information Needed\]
- **Language(s) (NLP):** English
- **License:** MIT License

### Dataset Sources \[optional\]

<!-- Provide the basic links for the dataset. -->

- **Repository:** https://github.com/kilian-group/phantom-wiki
- **Paper:** [PhantomWiki: On-Demand Datasets for Reasoning and Retrieval Evaluation](https://huggingface.co/papers/2502.20377)
- **Demo:** https://github.com/kilian-group/phantom-wiki/blob/main/demo.ipynb

## Uses

**We encourage users to generate a new (unique) PhantomWiki instance to combat data leakage and overfitting.**
PhantomWiki enables quantitative evaluation of the reasoning and retrieval capabilities of LLMs. See our full paper for analysis of frontier LLMs, including GPT-4o, Gemini-1.5-Flash, [Llama-3.3-70B-Instruct](https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct) and [DeepSeek-R1-Distill-Qwen-32B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B).

### Direct Use

PhantomWiki is intended to evaluate retrieval augmented generation (RAG) systems and agentic workflows.

### Out-of-Scope Use

<!-- This section addresses misuse, malicious use, and uses that the dataset will not work well for. -->

## Dataset Structure

PhantomWiki exposes three components, reflected in the 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 **split**.

## Dataset Creation

### Curation Rationale

Most mathematical and logical reasoning datasets do not explicity evaluate retrieval capabilities and
few retrieval datasets incorporate complex reasoning, save for a few exceptions (e.g., [BRIGHT](https://huggingface.co/datasets/xlangai/BRIGHT), [MultiHop-RAG](https://huggingface.co/datasets/yixuantt/MultiHopRAG)).
However, virtually all retrieval datasets are derived from Wikipedia or internet articles, which are contained in LLM training data.
We take the first steps toward a large-scale synthetic dataset that can evaluate LLMs' reasoning and retrieval capabilities.

### Source Data

This is a synthetic dataset. The extent to which we use real data is detailed as follows:

1. We sample surnames from among the most common surnames in the US population according to https://names.mongabay.com/most_common_surnames.htm
2. We sample first names using the `names` Python package (https://github.com/treyhunner/names). We thank the contributors for making this tool publicly available.
3. We sample jobs from the list of real-life jobs from the `faker` Python package. We thank the contributors for making this tool publicly available.
4. We sample hobbies from the list of real-life hobbies at https://www.kaggle.com/datasets/mrhell/list-of-hobbies. We are grateful to the original author for curating this list and making it publicly available.

#### Data Collection and Processing

This dataset was generated on commodity CPUs using Python and SWI-Prolog. See paper for full details of the generation pipeline, including timings.

#### Who are the source data producers?

<!-- This section describes the people or systems who originally created the data. It should also include self-reported demographic or identity information for the source data creators if this information is available. -->

N/A

### Annotations \[optional\]

<!-- If the dataset contains annotations which are not part of the initial data collection, use this section to describe them. -->

N/A

#### Annotation process

<!-- This section describes the annotation process such as annotation tools used in the process, the amount of data annotated, annotation guidelines provided to the annotators, interannotator statistics, annotation validation, etc. -->

N/A

#### Who are the annotators?

<!-- This section describes the people or systems who created the annotations. -->

N/A

#### Personal and Sensitive Information

<!-- State whether the dataset contains data that might be considered personal, sensitive, or private (e.g., data that reveals addresses, uniquely identifiable names or aliases, racial or ethnic origins, sexual orientations, religious beliefs, political opinions, financial or health data, etc.). If efforts were made to anonymize the data, describe the anonymization process. -->

PhantomWiki does not reference any personal or private data.

## Bias, Risks, and Limitations

<!-- This section is meant to convey both technical and sociotechnical limitations. -->

PhantomWiki generates large-scale corpora, reflecting fictional universes of characters and mimicking the style of fan-wiki websites. While sufficient for evaluating complex reasoning and retrieval capabilities of LLMs, PhantomWiki is limited to simplified family relations and attributes. Extending the complexity of PhantomWiki to the full scope of Wikipedia is an exciting future direction.

### Recommendations

<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->

PhantomWiki should be used as a benchmark to inform how LLMs should be used on reasoning- and retrieval-based tasks. For holistic evaluation on diverse tasks, PhantomWiki should be combined with other benchmarks.

## Citation \[optional\]

<!-- If there is a paper or blog post introducing the dataset, the APA and Bibtex information for that should go in this section. -->

**BibTeX:**

\[More Information Needed\]

**APA:**

\[More Information Needed\]

## Glossary \[optional\]

<!-- If relevant, include terms and calculations in this section that can help readers understand the dataset or dataset card. -->

\[More Information Needed\]

## More Information \[optional\]

\[More Information Needed\]

## Dataset Card Authors \[optional\]

\[More Information Needed\]

## Dataset Card Contact

agong@cs.cornell.edu