PaperlayerView exact revision on arXiv

Source-linked AI summary

Datasheets for Datasets

Timnit Gebru, Jamie Morgenstern, Briana Vecchione, Jennifer Wortman Vaughan, Hanna Wallach, Hal Daumé, Kate Crawford

arXiv:1803.09010v8cs.DBcs.AIcs.LG

TL;DR

Machine learning lacks a standardized process for documenting datasets, despite data's critical role in model behavior. The paper proposes dataset datasheets documenting lifecycle information and reports adoption by academic researchers and internal pilots at Microsoft, Google, and IBM.

  • Problem

    Despite data's critical role in model behavior, machine learning lacks a standardized process for documenting dataset creation and use.

    Source: §1 · p. 1

  • Method

    The paper proposes datasheets with lifecycle questions and a workflow for documenting datasets’ motivation, composition, collection, processing, uses, distribution, and maintenance.

    Source: §1 · pp. 1–2, §3 · p. 4

  • Results

    Academic researchers have adopted datasheets, and Microsoft, Google, and IBM have begun internal pilots.

    Source: §4 · p. 10

  • Takeaways & Limitations

    Datasheets offer creators an opportunity to improve communication with dataset consumers and signal their prioritization of transparency and accountability.

    Source: §4 · p. 11

    Datasheets are not a complete solution to mitigating societal biases or potential harms, since creators cannot anticipate every dataset use.

    Source: §4 · p. 11

Abstract

from arXiv · show

The machine learning community currently has no standardized process for documenting datasets, which can lead to severe consequences in high-stakes domains. To address this gap, we propose datasheets for datasets. In the electronics industry, every component, no matter how simple or complex, is accompanied with a datasheet that describes its operating characteristics, test results, recommended uses, and other information. By analogy, we propose that every dataset be accompanied with a datasheet that documents its motivation, composition, collection process, recommended uses, and so on. Datasheets for datasets will facilitate better communication between dataset creators and dataset consumers, and encourage the machine learning community to prioritize transparency and accountability.

1 Introduction

Dataset characteristics shape model behavior, yet machine learning lacks a standardized process for documenting datasets. The paper proposes datasheets to document datasets and improve transparency and accountability.

1 Introduction

  • Dataset mismatches and reflected societal biases can harm model performance and cause especially severe consequences in high-stakes domains.The paper notes that deployment context may differ from training or evaluation data, and that models can reproduce or amplify biases in datasets.Source: §1 · p. 1
  • Despite data provenance research in databases, machine learning has no standardized process for documenting dataset creation and use.Source: §1 · p. 1
  • The paper proposes datasheets documenting dataset motivation, composition, collection, and recommended uses to increase transparency and accountability.The proposal adapts the electronics-industry practice of accompanying components with datasheets describing characteristics, test results, and recommended usage.Source: §1 · pp. 1–2

1.1 Objectives

  • Datasheets aim to help creators reflect on dataset creation, distribution, and maintenance while giving consumers information to select suitable datasets and avoid misuse.Source: §1.1 · p. 2
  • Datasheets may also serve affected individuals and other stakeholders, and support reproducibility by helping researchers create alternative datasets with similar characteristics.Source: §1.1 · p. 2
  • The proposed questions are adaptable to domain and organizational workflows, and datasheet creation is intended to prompt reflection rather than be automated.The paper notes that questions suitable for public academic datasets may be less relevant to internal proprietary datasets.Source: §1.1 · pp. 2–3

2 Development Process

The authors developed the datasheet questions and workflow iteratively over roughly two years, testing drafts with dataset examples and product teams and incorporating feedback from researchers, practitioners, policy makers, and lawyers. This process reshaped the questions to better fit dataset lifecycle stages and elicit factual information without requiring legal judgments.

2 Development Process

  • Over roughly two years, the authors iteratively developed questions through dataset examples, product-team trials, broad public feedback, and legal review.They tested questions using example datasheets for Labeled Faces in the Wild and Pang and Lee’s polarity dataset [22], and circulated a draft that received comments from dozens of researchers, practitioners, and policy makers.Source: §2 · p. 3
  • Feedback led the authors to revise, reorder, and reword questions, add a “Uses” section, and integrate legal and ethical topics into lifecycle stages.Product teams were more likely to answer legal and ethical questions when they appeared alongside relevant lifecycle stages rather than in a separate section.Source: §2 · p. 3
  • After lawyers’ review, the authors removed explicit regulatory-compliance questions and added factual questions that avoid requiring creators to make legal judgments.They also deleted the separate “Legal and Ethical Considerations” section.Source: §2 · pp. 3–4

3 Questions and Workflow

The section proposes lifecycle-organized questions and a workflow for documenting datasets, prompting creators to reflect on creation, distribution, and maintenance; inapplicable questions may be skipped.

3.1 Motivation

  • Creators should document the dataset’s purpose, task or gap, creators, and funding interests.The questions are intended to clarify why the dataset was created and promote transparency about funding, especially for research datasets.Source: §3.1 · p. 4

3.2 Composition

  • Creators should document instance types, sampling and representativeness, contents and labels, splits, data quality and dependencies, plus confidentiality, identifiability, subpopulations, and sensitive data.The questions also address GDPR or comparable regulations and recommend a broad interpretation of whether a dataset relates to people, including text written by people.Source: §3.2 · pp. 5–6

3.3 Collection Process

  • Collection questions cover acquisition and validation, tools, sampling, workers and compensation, timing, ethics review, notice, consent and revocation, and impact analyses.Creators should review the questions before collection to flag potential issues, then answer them afterward; the information can also help others create alternative datasets with similar characteristics.Source: §3.3 · pp. 6–8

3.4 Preprocessing/cleaning/labeling

  • Creators should document preprocessing, cleaning, and labeling, preserve raw data when available, and identify the software so consumers can judge task compatibility.The questions are to be reviewed before these tasks and answered once they are complete; for example, bag-of-words text is unsuitable for tasks involving word order.Source: §3.4 · p. 8

3.5 Uses

  • Creators should describe prior and potential tasks, risks or harms and possible mitigations, and tasks for which the dataset should not be used.These disclosures are intended to help consumers make informed decisions and avoid risks or harms.Source: §3.5 · pp. 8–9

3.6 Distribution

  • Before distribution, creators should report recipients, channel and timing, DOI, licenses and fees, third-party restrictions, and applicable export or regulatory controls.The questions cover both distribution within the creating entity and distribution to outside parties.Source: §3.6 · p. 9

3.7 Maintenance

  • Creators should identify maintainers and contact routes, errata and update or version plans, retention limits, and processes for contributions and validation.The questions encourage creators to plan dataset maintenance and communicate that plan to consumers before distribution.Source: §3.7 · pp. 9–10

4 Impact and Challenges

Datasheets have gained early adoption, but wider use faces workflow, dynamic-data, societal-risk, expertise, and overhead challenges; the authors argue their communication and accountability benefits outweigh the costs.

Impact and Challenges

  • Datasheets gained traction through academic releases and corporate pilots, alongside related model cards, data cards, factsheets, nutrition labels, and industry-wide guidance.Examples include academic datasets [e.g., 7, 10, 23, 26], model cards [20], the Open Images data card, IBM factsheets, and the Dataset Nutrition Label; Microsoft, Google, and IBM piloted datasheets internally, while the Partnership on AI developed industry guidance.Source: §4 · pp. 10–11
  • Wider adoption may require adapting questions and workflows to organizational infrastructure, while infrequently changing datasets should receive updated datasheets when revised.The questions and workflow may pose particular problems for dynamic datasets.Source: §4 · p. 11
  • Datasheets cannot fully mitigate societal bias or harm because creators cannot anticipate every use, and demographic labels may be unavailable for privacy or data-protection reasons.Datasheets about people may also require collaboration with anthropology, sociology, or science and technology studies experts to account for contextual social, historical, and geographical factors in respectful data collection.Source: §4 · p. 11
  • Creating datasheets imposes unavoidable time and organizational costs, even if they reduce one-off questions, and may require changes to workflows and incentives.Source: §4 · p. 11
  • Datasheets improve creator-consumer communication and let creators demonstrate transparency and accountability; the authors believe these benefits outweigh the costs.Source: §4 · p. 11

A Appendix

The appendix illustrates a datasheet for Pang and Lee’s polarity dataset [22], documenting its composition, preprocessing, and use-related limitations.

Composition

  • The dataset comprises movie reviews from newsgroup postings with binary positive/negative sentiment labels, and contains 1,400 instances in v1.x or 2,000 in v2.0.The labels distinguish strongly positive from strongly negative reviews.Source: A Appendix · p. 15

Preprocessing/cleaning/labeling

  • Preprocessing retained strongly rated reviews with explicit ratings, automatically extracted star-rating labels, discarded middle ratings, capped reviews at 20 per author per label, and later removed non-English reviews.Half-star ratings were occasionally missed, affecting negative-example labels.Source: A Appendix · p. 17

Other uses

  • The dataset may also support movie-review modeling, sentiment-lexicon induction, or automatic review generation.Source: A Appendix · p. 17

Use limitations

  • Because the data covers only movie reviews, systems trained on it may not generalize to other sentiment tasks and should not inform consequential decisions without additional verification.The datasheet says the preprocessed data had minimal harm risk because it was already public and names and email addresses were removed.Source: A Appendix · p. 17

Related papers

Loading 1803.09010v8…