Addressing "Documentation Debt" in Machine Learning Research: A Retrospective Datasheet for BookCorpus

2021/05/11 by Jack Bandy, Nicholas Vincent, Bandy, Jack +1 · 2 voices · 35 citations
Computer Science · Decision Sciences · #Accounting #Business #Computer science #Data Quality and Management #Datasheet #Debt #Documentation #Explainable Artificial Intelligence (XAI) #Finance #Operating system #Programming language #Topic Modeling #cs.CL #cs.CY #cs.LG

paper · pdf · doi:10.48550/arxiv.2105.05241

published in arXiv (Cornell University) (Cornell University) · Working paper

arxiv created 2021/05/11 · openalex publication_date 2021/05/11 · arxiv updated 2021/05/12 · openalex created_date 2021/05/24 · openalex updated_date 2026/07/28

Abstract

Recent literature has underscored the importance of dataset documentation work for machine learning, and part of this work involves addressing "documentation debt" for datasets that have been used widely but documented sparsely. This paper aims to help address documentation debt for BookCorpus, a popular text dataset for training large language models. Notably, researchers have used BookCorpus to train OpenAI's GPT-N models and Google's BERT models, even though little to no documentation exists about the dataset's motivation, composition, collection process, etc. We offer a preliminary datasheet that provides key context and information about BookCorpus, highlighting several notable deficiencies. In particular, we find evidence that (1) BookCorpus likely violates copyright restrictions for many books, (2) BookCorpus contains thousands of duplicated books, and (3) BookCorpus exhibits significant skews in genre representation. We also find hints of other potential deficiencies that call for future research, including problematic content, potential skews in religious representation, and lopsided author contributions. While more work remains, this initial effort to provide a datasheet for BookCorpus adds to growing literature that urges more careful and systematic documentation for machine learning datasets.

Citations

• On the Dangers of Stochastic Parrots
• Datasheets for Datasets

Cited by

• A Human-Centric Framework for Data Attribution in Large Language Models
• Provocations from the Humanities for Generative AI Research
• On the Opportunities and Risks of Foundation Models
• The BigScience ROOTS Corpus: A 1.6TB Composite Multilingual Dataset
• AI-generated stories favour stability over change: homogeneity and cultural stereotyping in narratives generated by gpt-4o-mini
• Uncertainty in Semantic Language Modeling with PIXELS
• Frequency Effects on Syntactic Rule Learning in Transformers
• Biased Local SGD for Efficient Deep Learning on Heterogeneous Systems
• The value of books in the age of generative AI training data
• From Reflection to Repair: A Scoping Review of Dataset Documentation Tools
• Certified Mitigation of Worst-Case LLM Copyright Infringement
• BookCorpus [wikipedia]

Discussions

• The news that AI was being trained on stolen works? It was broken first about the BooksCorpus, on May 11 of 2021, and it was discussed in major news media thereafter. Romance authors talked about it. [bsky, 1265 points, 3 comments]
• But again, that doesn't often happen, especially for data used by companies selling large models. Let's go back to books. Could we work backwards and create a "datasheet" for one of the books datasets [bsky, 4 points, 1 comments]

Related

• Using Large Language Models to Enrich the Documentation of Datasets for Machine Learning
• Institutional Books 1.0: A 242B token dataset from Harvard Library's collections, refined for accuracy and usability
• The value of books in the age of generative AI training data
• The Data Provenance Initiative: A Large Scale Audit of Dataset Licensing & Attribution in AI
• Digger: Detecting Copyright Content Mis-usage in Large Language Model Training