Dataset Summary

RetrievalQA is a short-form open-domain question answering (QA) dataset consisting of 1,271 questions covering new world and long-tail knowledge. We ensure the knowledge necessary to answer the questions is absent from most LLMs. Therefore, LLMs must truthfully decide whether to retrieve to be able to answer the questions correctly.

RetrievalQA enables us to evaluate the effectiveness of adaptive retrieval-augmented generation (RAG) approaches, an aspect predominantly overlooked in prior studies and recent RAG evaluation systems, which focus only on task performance, the relevance of retrieval context or the faithfulness of answers.

Dataset Sources

Repository: https://github.com/hyintell/RetrievalQA
Paper: https://arxiv.org/abs/2402.16457

Dataset Structure

Here is an example of a data instance:

{
  "data_source": "realtimeqa", 
  "question_id": "realtimeqa_20231013_1", 
  "question": "What percentage of couples are 'sleep divorced', according to new research?", 
  "ground_truth": ["15%"], 
  "context": [
    {
      "title": "Do We Sleep Longer When We Share a Bed?", 
      "text": "1.4% of respondents have started a sleep divorce, or sleeping separately from their partner, and maintained it in the past year. Adults who have ..."
    }, ...
  ]
}

where:

data_source: the origin dataset of the question comes from
question: the question
ground_truth: a list of possible answers
context: a list of dictionaries of retrieved relevant evidence. Note that the title of the document might be empty.

Citation

@misc{zhang2024retrievalqa,
      title={RetrievalQA: Assessing Adaptive Retrieval-Augmented Generation for Short-form Open-Domain Question Answering}, 
      author={Zihan Zhang and Meng Fang and Ling Chen},
      year={2024},
      eprint={2402.16457},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}