wentingzhao/WildHallucinations

WildHallucinations is designed for evaluating the factuality of LLMs. Its core idea is to prompt LLMs to generate and fact-check information about a diverse set of entities. WildHallucinations consists of 7917 entities extracted from WildChat and a knowledge source. These entities come from English conversations that are marked as non-toxic. As described in the main paper, we apply extensive filtering for quality control, especially for removing entities with more than one meaning. The knowledge source is constructed from Google search API. We scrape the top 10 web pages for each entity. Additional cleaning process can be found in the paper.

To use the dataset:

from datasets import load_dataset
ds = load_dataset("wentingzhao/WildHallucinations", split="train")

Dataset Columns:

• entity (string): the entity name
• perplexity (float): the perplexity of the entity measured by the Llama-3-8B model
• info (string): the web information about the entity scraped from Google search results
• category (string): the category of the entity annotated by either an author or GPT-4o
• wiki (Boolean): whether any information about the entity comes from wikipedia.org

Citation Information

Please consider citing our paper if you find this dataset useful:

@article{
    zhao2024wildhallucinations,
    title={WildHallucinations: Evaluating Long-form Factuality in LLMs with Real-World Entity Queries},
    author={Wenting Zhao, Tanya Goyal, Yu Ying Chiu, Liwei Jiang, Benjamin Newman, Abhilasha Ravichander, Khyathi Chandu, Ronan Le Bras, Claire Cardie, Yuntian Deng, Yejin Choi},
    journal={arXiv preprint arXiv:2407.17468},
    year={2024}
}