---

license: apache-2.0 task_categories: - question-answering size_categories: - 10K<n<100K

Dataset Card for NovelQA

NovelQA is a benchmark for testing the long-text ability of LLMs.

The Dataset Viewer is automatically generated and only display the novel content. We do not know how to remove or remake it, so you can just ignore it.

Dataset Details

Dataset Sources

• Repository: https://github.com/NovelQA/novelqa.github.io
• Leaderboard: https://novelqa.github.io
• Competition:: https://www.codabench.org/competitions/2295/
• Paper: https://arxiv.org/abs/2403.12766

Uses

Directly downloading

You can directly download the NovelQA.zip, which contains all files of Raw_Novels/, Data/ and Demonstration/.

Through API

You might also use this dataset through the Huggingface dataset package as follows.

from datasets import load_dataset
dataset = load_dataset("NovelQA/NovelQA", data_files = {
    "book": "Raw_Novels/*.txt",
    "ques": "Data/*.json"
}, streaming=True)

books = dataset["book"]
ques = dataset["ques"]

Dataset Structure

The dataset is structured as follows.

- NovelQA.zip                  // This zip file contains all of the novels, data and demonstration 
 
- NovelQA
  | - book                     // the book contents of the novels
  |   | - booktitle1.txt
  |   | - booktitle2.txt
  |   | - ...
  |
  | - ques                     // the corresponding QA-pairs of each book
      | - booktitle1.json
      | - booktitle2.json
      | - ...

Among the json files, each file includes a list of dicts, each of which is structured as follows.

    {
        "Question": "The input question",
        "Options": [
            "Option A",
            "Option B",
            "Option C",
            "Option D"
        ],
        "Complex": "A complexity level among mh, sh, and dtl",
        "Aspect": "An aspect from times, meaning, span, settg, relat, character, and plot"
    },

Citation

BibTeX:

@misc{wang2024novelqa,
      title={NovelQA: A Benchmark for Long-Range Novel Question Answering}, 
      author={Cunxiang Wang and Ruoxi Ning and Boqi Pan and Tonghui Wu and Qipeng Guo and Cheng Deng and Guangsheng Bao and Qian Wang and Yue Zhang},
      year={2024},
      eprint={2403.12766},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

Term

Your participation and submission to this benchmark will naturally give your consent to the following terms.

The input data are only for internal evaluation use. Please do not publicly spread the input data online. The competition hosts are not responsible for any possible violation of novel copyright caused by the participants' spreading the input data publicly online.

Contact

If you find problems downloading or using this dataset, please contact the first authors from the Arxiv paper to get access to the dataset.