Requesting access confirms that you have read this card and understand that the gate does not modify the upstream MIT license or imply endorsement by the original authors.

EntityQuestions

This repository is an unofficial research redistribution of the EntityQuestions dataset created by Christopher Sciavolino, Zexuan Zhong, Jinhyuk Lee, and Danqi Chen at Princeton NLP.

It is provided for convenient use through the Hugging Face Hub. This mirror is not maintained, sponsored, or endorsed by the original authors.

Dataset summary

EntityQuestions is an English entity-centric question-answering benchmark introduced in Simple Entity-Centric Questions Challenge Dense Retrievers. Its questions are based on facts derived from T-REx and Wikidata and were designed to evaluate retrieval and generalization across entities and question patterns.

Original sources:

Dataset and code: https://github.com/princeton-nlp/EntityQuestions
Paper: https://aclanthology.org/2021.emnlp-main.496/
arXiv: https://arxiv.org/abs/2109.08535

Data structure

Each JSONL record contains:

{
  "question": "What kind of work does Vidar Kleppe do?",
  "answers": ["politician"],
  "relation": "P106"
}

question: an English entity-centric question.
answers: one or more acceptable answer strings.
relation: the Wikidata property identifier recovered from the upstream source filename during repackaging.

Splits

Split	Examples
Train	176,560
Validation	22,068
Test	22,075
Total	220,703

The upstream dev split is exposed as validation in this redistribution. The unmodified upstream archive is also included as dataset_original.zip (SHA-256: 015eb64a12fa2f9b2f95d2c78bba69779e0d200dbbd4b0772f3ce57bc00da9c5).

Important: no SliCK annotations

This repository contains the original EntityQuestions question-answer pairs. It does not contain the annotations generated in Does Fine-Tuning LLMs on New Knowledge Encourage Hallucinations? (HighlyKnown, MaybeKnown, WeaklyKnown, Unknown), nor its sampled model outputs or estimated P_correct values.

Provenance and attribution

Authorship remains with the original Princeton NLP researchers. According to the upstream repository, EntityQuestions derives from T-REx and Wikidata.

This mirror should not be cited as the original source. Please cite the original paper:

@inproceedings{sciavolino2021simple,
  title     = {Simple Entity-centric Questions Challenge Dense Retrievers},
  author    = {Sciavolino, Christopher and Zhong, Zexuan and
               Lee, Jinhyuk and Chen, Danqi},
  booktitle = {Empirical Methods in Natural Language Processing (EMNLP)},
  year      = {2021}
}

License and gated access

The original authors release the dataset and code under the MIT License. Their repository states that the source resources are T-REx, released under MIT, and Wikidata, released under CC0 1.0.

The automatic access gate on this mirror is used only to record access requests and contact information. It does not change those licenses, impose additional use restrictions, make the data exclusive, or imply endorsement by the original authors. Preserve the original copyright and license notices when redistributing the data.

Limitations

The data is derived from a particular snapshot of knowledge sources and may contain outdated facts, incomplete aliases, ambiguous answers, or template artifacts. Exact-match evaluation may undercount semantically correct answers. Users should independently assess suitability, licensing, privacy, and benchmark-contamination risks for their application.

Downloads last month: 27

Total file size:

27.1 MB

Paper for costadev00/EntityQuestions

Simple Entity-Centric Questions Challenge Dense Retrievers

Paper • 2109.08535 • Published Sep 17, 2021