vincentkoc/hover-parquet

Dataset Card for HoVer (Parquet Format)

Note: This is a scriptless, Parquet-based version of the HoVer dataset for seamless integration with HuggingFace datasets library. No trust_remote_code required!

Quick Start

from datasets import load_dataset

# Load the dataset (no trust_remote_code needed!)
dataset = load_dataset("vincentkoc/hover-parquet")

# Access splits
train = dataset["train"]
validation = dataset["validation"]
test = dataset["test"]

# Example usage
print(train[0])
# {
#   'id': 0,
#   'uid': '330ca632-e83f-4011-b11b-0d0158145036',
#   'claim': 'Skagen Painter Peder Severin Krøyer favored naturalism...',
#   'supporting_facts': [{'key': 'Kristian Zahrtmann', 'value': 0}, ...],
#   'label': 1,  # 0: NOT_SUPPORTED, 1: SUPPORTED
#   'num_hops': 3,
#   'hpqa_id': '5ab7a86d5542995dae37e986'
# }

Dataset Summary

HoVer (HOP VERification) is an open-domain, many-hop fact extraction and claim verification dataset built upon the Wikipedia corpus. The dataset contains claims that require reasoning over multiple documents (multi-hop) to verify whether they are supported or not supported by evidence.

The original 2-hop claims are adapted from question-answer pairs from HotpotQA. It was collected by a team of NLP researchers at UNC Chapel Hill and Verisk Analytics.

This version provides the dataset in Parquet format for efficient loading and compatibility with modern data processing pipelines, eliminating the need for custom loading scripts.

Supported Tasks and Leaderboards

• Fact Verification: Determine whether a claim is SUPPORTED or NOT_SUPPORTED based on evidence from Wikipedia articles
• Multi-hop Reasoning: Claims require reasoning across multiple documents (indicated by num_hops field)
• Evidence Retrieval: Identify relevant supporting facts from source documents

The official leaderboard is available at https://hover-nlp.github.io/

Languages

English (en)

Dataset Structure

Data Instances

A sample training set example:

{
  "id": 14856,
  "uid": "a0cf45ea-b5cd-4c4e-9ffa-73b39ebd78ce",
  "claim": "The park at which Tivolis Koncertsal is located opened on 15 August 1843.",
  "supporting_facts": [
    {"key": "Tivolis Koncertsal", "value": 0},
    {"key": "Tivoli Gardens", "value": 1}
  ],
  "label": 1,
  "num_hops": 2,
  "hpqa_id": "5abca1a55542993a06baf937"
}

Note: In the test set, only id, uid, and claim fields contain meaningful data. The label is set to -1, num_hops to -1, hpqa_id to "None", and supporting_facts is an empty list, as these are withheld for evaluation purposes.

Data Fields

• id (int32): Sequential identifier for the example within its split
• uid (string): Unique identifier (UUID) for the claim
• claim (string): The claim statement to be verified
• supporting_facts (list): List of evidence facts, where each fact contains:
  • key (string): Title of the Wikipedia article
  • value (int32): Sentence index within that article
• label (ClassLabel): Verification label with values:
  • 0: NOT_SUPPORTED - The claim is not supported by the evidence
  • 1: SUPPORTED - The claim is supported by the evidence
  • -1: Unknown (used in test set)
• num_hops (int32): Number of reasoning hops required (typically 2-4 for this dataset)
• hpqa_id (string): Original HotpotQA question ID from which the claim was derived

Data Splits

Split	Examples
Train	18,171
Validation	4,000
Test	4,000
Total	26,171

The splits maintain the original distribution from the HoVer dataset.

Dataset Creation

Curation Rationale

HoVer was created to address the challenge of multi-hop fact verification, where claims require reasoning across multiple documents. The dataset was built to push the boundaries of claim verification systems beyond single-document fact-checking.

Source Data

The dataset is built upon Wikipedia as the knowledge source. Claims are adapted from HotpotQA question-answer pairs and modified to create verification statements that require multi-hop reasoning.

Annotations

The dataset was annotated by expert annotators who identified supporting facts across multiple Wikipedia articles and determined whether claims were supported or not supported by the evidence.

Additional Information

Licensing Information

This dataset is licensed under the MIT License.

Citation Information

@inproceedings{jiang2020hover,
  title={{HoVer}: A Dataset for Many-Hop Fact Extraction And Claim Verification},
  author={Yichen Jiang and Shikha Bordia and Zheng Zhong and Charles Dognin and Maneesh Singh and Mohit Bansal},
  booktitle={Findings of the Conference on Empirical Methods in Natural Language Processing (EMNLP)},
  year={2020}
}

Contributions

Thanks to @abhishekkrthakur for adding the original dataset and @vincentkoc for creating this Parquet version.