Datasets:
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]
- Repository: https://github.com/kilian-group/phantom-wiki
- Paper: PhantomWiki: On-Demand Datasets for Reasoning and Retrieval Evaluation
- 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 and DeepSeek-R1-Distill-Qwen-32B.
Direct Use
PhantomWiki is intended to evaluate retrieval augmented generation (RAG) systems and agentic workflows.
Out-of-Scope Use
Dataset Structure
PhantomWiki exposes three components, reflected in the three configurations:
question-answer: Question-answer pairs generated using a context-free grammartext-corpus: Documents generated using natural-language templatesdatabase: 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, MultiHop-RAG). 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:
- We sample surnames from among the most common surnames in the US population according to https://names.mongabay.com/most_common_surnames.htm
- We sample first names using the
namesPython package (https://github.com/treyhunner/names). We thank the contributors for making this tool publicly available. - We sample jobs from the list of real-life jobs from the
fakerPython package. We thank the contributors for making this tool publicly available. - 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?
N/A
Annotations [optional]
N/A
Annotation process
N/A
Who are the annotators?
N/A
Personal and Sensitive Information
PhantomWiki does not reference any personal or private data.
Bias, Risks, and 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
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]
BibTeX:
[More Information Needed]
APA:
[More Information Needed]
Glossary [optional]
[More Information Needed]
More Information [optional]
[More Information Needed]
Dataset Card Authors [optional]
[More Information Needed]