BookCorpus Datasheet

Documentation effort for the BookCorpus dataset. (Link to arXiv paper)

Summary Data Card

If you like the nutrition label style, here is an Overleaf Template you can use to recreate and modify it.

Context

This datasheet is inspired by Bender and Gebru et al.'s notion of "documentation debt," and our observation that BookCorpus has been used widely but documented sparsely.

We plan to incorporate these findings in the Hugging Face entry for BookCorpus.

The Data

BookCorpus dataset

BookCorpus is no longer available from the original authors, though we obtained the dataset directly from their website through a security vulnerability which we have since notified them about.

We are not redistributing BookCorpus in full, however we include the following metadata files in the data/BookCorpus folder:

• books_in_bookcorpus.csv lists all text files (books) in BookCorpus, as downloaded from the authors' website, including the book's location, category, word_count, and disk_usage
• sentences_counted_8+.csv lists all sentences that occurred eight or more times
• count_of_sentence_counts.csv lists data about the number of unique_sentences (including a random_sentence) that occurred n times
• stolen_books.md (pretty) lists books included in the original BookCorpus that now cost money to download from Smashwords.com

BookCorpusOpen dataset

"BookCorpusOpen" is included as "BookCorpus2" in the Pile. Here are some background details about how BookCorpusOpen (also referred to as OpenBookCorpus, Books1, and BookCorpusNew) was constructed and published. The data/BookCorpusOpen folder contains one file:

• 2020-08-27-epub_urls.txt lists smashwords.com URLs used to collect BookCorpusOpen

Smashwords21 dataset

We collected all books listed on Smashwords as of April 2021. The data/Smashwords21 folder contains two files:

• smashwords_april_2021.csv is the direct output from our scraping program that includes each book's Link, Title, Author,Price,Words, and more
• smashwords_april_2021_dedup.csv is a deduplicated version of the above data (based on each book's Link), which we use in our analysis

Other Dataset recreation efforts

• Replicate Toronto BookCorpus
• Homemade BookCorpus

The Code

We include three files used in our analysis:

• DataSheet contains the code we used for most of our analysis
• Sentence Analysis contains the code we used to inspect sentences
• WordsAndBooksPerAuthor contains the code we used to analyze author contributions

How to Cite

To cite the original "datasheets for datasets" paper:

@inproceedings{gebru2018datasheets,
  title={Datasheets for datasets},
  author={Gebru, Timnit and Morgenstern, Jamie and Vecchione, Briana and Vaughan, Jennifer Wortman and Wallach, Hanna and Daum{\'e} III, Hal and Crawford, Kate},
  booktitle={Proceedings of the 5th Workshop on Fairness, Accountability, and Transparency in Machine Learning},
  url={https://arxiv.org/abs/1803.09010},
  year={2018}
}

To cite our paper (i.e. this BookCorpus datasheet / data card):

@article{bandy2021addressing,
      title={Addressing "Documentation Debt" in Machine Learning Research: A Retrospective Datasheet for BookCorpus},
      author={Bandy, Jack and Vincent, Nicholas},
      year={2021},
      journal={arXiv preprint arXiv:2105.05241},
      url={https://arxiv.org/abs/2105.05241}
}

---

💬 Head to the discussion with any questions or comments, or reach out directly to Jack and/or Nick!