Data plays a critical role in machine learning. Every machine learning model is trained and evaluated using data, quite often in the form of static datasets. The characteristics of these datasets fundamentally influence a model's behavior: a model is unlikely to perform well in the wild if its deployment context does not match its training or evaluation datasets, or if these datasets reflect unwanted societal biases. Mismatches like this can have especially severe consequences when machine learning models are used in high-stakes domains, such as criminal justice,1,13,24 hiring,19 critical infrastructure,11,21 and finance.18 Even in other domains, mismatches may lead to loss of revenue or public relations setbacks. Of particular concern are recent examples showing that machine learning models can reproduce or amplify unwanted societal biases reflected in training datasets.4,5,12 For these and other reasons, the World Economic Forum suggests all entities should document the provenance, creation, and use of machine learning datasets to avoid discriminatory outcomes.25
Although data provenance has been studied extensively in the databases community,3,8 it is rarely discussed in the machine learning community. Documenting the creation and use of datasets has received even less attention. Despite the importance of data to machine learning, there is currently no standardized process for documenting machine learning datasets.
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 describing its operating characteristics, test results, recommended usage, 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 have the potential to increase transparency and accountability within the machine learning community, mitigate unwanted societal biases in machine learning models, facilitate greater reproducibility of machine learning results, and help researchers and practitioners to select more appropriate datasets for their chosen tasks.
After outlining our objectives, we describe the process by which we developed datasheets for datasets. We then provide a set of questions designed to elicit the information that a datasheet for a dataset might contain, as well as a workflow for dataset creators to use when answering these questions. We conclude with a summary of the impact to date of datasheets for datasets and a discussion of implementation challenges and avenues for future work.
Objectives. Datasheets for datasets are intended to address the needs of two key stakeholder groups: dataset creators and dataset consumers. For dataset creators, the primary objective is to encourage careful reflection on the process of creating, distributing, and maintaining a dataset, including any underlying assumptions, potential risks or harms, and implications of use. For dataset consumers, the primary objective is to ensure they have the information they need to make informed decisions about using a dataset. Transparency on the part of dataset creators is necessary for dataset consumers to be sufficiently well informed that they can select appropriate datasets for their chosen tasks and avoid unintentional misuse.a
Beyond these two key stakeholder groups, datasheets for datasets may be valuable to policy makers, consumer advocates, investigative journalists, individuals whose data is included in datasets, and individuals who may be impacted by models trained or evaluated using datasets. They also serve a secondary objective of facilitating greater reproducibility of machine learning results: researchers and practitioners without access to a dataset may be able to use the information in its datasheet to create alternative datasets with similar characteristics.
Although we provide a set of questions designed to elicit the information a datasheet for a dataset might contain, these questions are not intended to be prescriptive. Indeed, we expect that datasheets will necessarily vary depending on factors such as the domain or existing organizational infrastructure and workflows. For example, some the questions are appropriate for academic researchers publicly releasing datasets for the purpose of enabling future research, but less relevant for product teams creating internal datasets for training proprietary models. As another example, Bender and Friedman2 outline a proposal similar to datasheets for datasets specifically intended for language-based datasets. Their questions may be naturally integrated into a datasheet for a language-based dataset as appropriate.
We emphasize that the process of creating a datasheet is not intended to be automated. Although automated documentation processes are convenient, they run counter to our objective of encouraging dataset creators to carefully reflect on the process of creating, distributing, and maintaining a dataset.
Here, we refined the questions and workflow provided over a period of approximately two years, incorporating many rounds of feedback.
First, leveraging our own experiences as researchers with diverse backgrounds working in different domains and institutions, we drew on our knowledge of dataset characteristics, unintentional misuse, unwanted societal biases, and other issues to produce an initial set of questions designed to elicit information about these topics. We then "tested" these questions by creating example datasheets for two widely used datasets: Labeled Faces in the Wild16 and Pang and Lee's polarity dataset.22 We chose these datasets in large part because their creators provided exemplary documentation, allowing us to easily find the answers to many of the questions. While creating these example datasheets, we found gaps in the questions, as well as redundancies and lack of clarity. We therefore refined the questions and distributed them to product teams in two major U.S.-based technology companies, in some cases helping teams to create datasheets for their datasets and observing where the questions did not achieve their intended objectives. Contemporaneously, we circulated an initial draft of this article to colleagues through social media and on arXiv (draft posted Mar. 23, 2018). Via these channels we received extensive comments from dozens of researchers, practitioners, and policy makers. We also worked with a team of lawyers to review the questions from a legal perspective.
We incorporated this feedback to yield the questions and workflow provided in the next section: We added and removed questions, refined the content of the questions, and reordered the questions to better match the key stages of the dataset life cycle. Based on our experiences with product teams, we reworded the questions to discourage yes/no answers, added a section on "Uses," and deleted a section on "Legal and Ethical Considerations." We found that product teams were more likely to answer questions about legal and ethical considerations if they were integrated into sections about the relevant stages of the dataset lifecycle rather than grouped together. Finally, following feedback from the team of lawyers, we removed questions that explicitly asked about compliance with regulations, and introduced factual questions intended to elicit relevant information about compliance without requiring dataset creators to make legal judgments.
In this section, we provide a set of questions designed to elicit the information that a datasheet for a dataset might contain, as well as a workflow for dataset creators to use when answering these questions. The questions are grouped into sections that approximately match the key stages of the dataset lifecycle: motivation, composition, collection process, preprocessing/cleaning/labeling, uses, distribution, and maintenance. This grouping encourages dataset creators to reflect on the process of creating, distributing, and maintaining a dataset, and even alter this process in response to their reflection. We note that not all questions will be applicable to all datasets; those that do not apply should be skipped.
To illustrate how these questions might be answered in practice, we produced an appendix that includes an example datasheet for Pang and Lee's polarity dataset.22 (The appendix is available online at https://dl.acm.org/doi/10.1145/3458723.) We answered some of the questions with "Unknown to the authors of the datasheet." This is because we did not create the dataset ourselves and could not find the answers to these questions in the available documentation. For an example of a datasheet that was created by the creators of the corresponding dataset, please see that of Cao and Daumé.6,b We note that even dataset creators may be unable to answer all the questions provided here. We recommend answering as many questions as possible rather than skipping the datasheet creation process entirely.
Motivation. The following questions are primarily intended to encourage dataset creators to clearly articulate their reasons for creating the dataset and to promote transparency about funding interests. The latter may be particularly relevant for datasets created for research purposes.
Composition. Dataset creators should read through these questions prior to any data collection and then provide answers once data collection is complete. Most of the questions here are intended to provide dataset consumers with the information they need to make informed decisions about using the dataset for their chosen tasks. Some of the questions are designed to elicit information about compliance with the EU's General Data Protection Regulation (GDPR) or comparable regulations in other jurisdictions.
Questions that apply only to datasets that relate to people are grouped together at the end of the section. We recommend taking a broad interpretation of whether a dataset relates to people. For example, any dataset containing text that was written by people relates to people.
Datasheets for datasets have the potential to increase transparency and accountability within the ML community, mitigate unwanted societal biases in ML models, facilitate greater reproducibility of ML results, and help researchers and practitioners select more appropriate datasets for their chosen tasks.
If the dataset does not relate to people, you may skip the remaining questions in this section.
Collection process. As with the questions in the previous section, dataset creators should read through these questions prior to any data collection to flag potential issues and then provide answers once collection is complete. In addition to the goals outlined earlier, the following questions are designed to elicit information that may help researchers and practitioners to create alternative datasets with similar characteristics. Again, questions that apply only to datasets that relate to people are grouped together at the end of the section.
The process of creating a datasheet is not intended to be automated. Although automated documentation processes are convenient, they run counter to our objective of encouraging dataset creators to carefully reflect on the process of creating, distributing, and maintaining a dataset.
If the dataset does not relate to people, you may skip the remaining questions in this section.
Preprocessing/cleaning/labeling. Dataset creators should read through these questions prior to any preprocessing, cleaning, or labeling and then provide answers once these tasks are complete. The questions in this section are intended to provide dataset consumers with the information they need to determine whether the "raw" data has been processed in ways that are compatible with their chosen tasks. For example, text that has been converted into a "bag-of-words" is not suitable for tasks involving word order.
Uses. The following questions are intended to encourage dataset creators to reflect on the tasks for which the dataset should and should not be used. By explicitly highlighting these tasks, dataset creators can help dataset consumers to make informed decisions, thereby avoiding potential risks or harms.
Distribution. Dataset creators should provide answers to these questions prior to distributing the dataset either internally within the entity on behalf of which the dataset was created or externally to third parties.
Maintenance. As with the previous questions, dataset creators should provide answers to these questions prior to distributing the dataset. The questions in this section are intended to encourage dataset creators to plan for dataset maintenance and communicate this plan to dataset consumers.
Since circulating an initial draft of this article in March 2018, datasheets for datasets have already gained traction in a number of settings. Academic researchers have adopted our proposal and released datasets with accompanying datasheets.7,10,23,26 Microsoft, Google, and IBM have begun to pilot datasheets for datasets internally within product teams. Researchers at Google published follow-up work on model cards that document machine learning models20 and released a data card (a lightweight version of a datasheet) along with the Open Images dataset.17 Researchers at IBM proposed factsheets14 that document various characteristics of AI services, including whether the datasets used to develop the services are accompanied with datasheets. The Data Nutrition Project incorporated some of the questions provided in the previous section into the latest release of their Dataset Nutrition Label.9 Finally, the Partnership on AI, a multi-stakeholder organization focused on studying and formulating best practices for developing and deploying AI technologies, is working on industry-wide documentation guidance that builds on datasheets for datasets, model cards, and factsheets.c
These initial successes have also revealed implementation challenges that may need to be addressed to support wider adoption. Chief among them is the need for dataset creators to modify the questions and workflow provided earlier based on their existing organizational infrastructure and workflows. We also note that the questions and workflow may pose problems for dynamic datasets. If a dataset changes only infrequently, we recommend accompanying updated versions with updated datasheets.
Datasheets for datasets do not provide a complete solution to mitigating unwanted societal biases or potential risks or harms. Dataset creators cannot anticipate every possible use of a dataset, and identifying unwanted societal biases often requires additional labels indicating demographic information about individuals, which may not be available to dataset creators for reasons including those individuals' data protection and privacy.15
When creating datasets that relate to people, and hence their accompanying datasheets, it may be necessary for dataset creators to work with experts in other domains such as anthropology, sociology, and science and technology studies. There are complex and contextual social, historical, and geographical factors that influence how best to collect data from individuals in a manner that is respectful.
Finally, creating datasheets for datasets will necessarily impose overhead on dataset creators. Although datasheets may reduce the amount of time that dataset creators spend answering one-off questions about datasets, the process of creating a datasheet will always take time, and organizational infrastructure and workflows—not to mention incentives—will need to be modified to accommodate this investment.
Despite these implementation challenges, there are many benefits to creating datasheets for datasets. In addition to facilitating better communication between dataset creators and dataset consumers, datasheets provide an opportunity for dataset creators to distinguish themselves as prioritizing transparency and accountability. Ultimately, we believe that the benefits to the machine learning community outweigh the costs.
We thank P. Bailey, E. Bender, Y. Bengio, S. Bird, S. Brown, S. Bowles, J. Buolamwini, A. Casari, E. Charran, A. Couillault, L. Dauterman, L. Dodds, M. Dudík, M. Ekstrand, N. Elhadad, M. Golebiewski, N. Gonsalves, M. Hansen, A. Hickl, M. Hoffman, S. Hoogerwerf, E. Horvitz, M. Huang, S. Kallumadi, E. Kamar, K. Kenthapadi, E. Kiciman, J. Krones, E. Learned-Miller, L. Lee, J. Leidner, R. Mauceri, B. Mcfee, E. McReynolds, B. Micu, M. Mitchell, S. Mudnal, B. O'Connor, T. Padilla, B. Pang, A. Parikh, L. Peets, A. Perina, M. Philips, B. Place, S. Rao, J. Ren, D. Van Riper, A. Roth, C. Rudin, B. Shneiderman, B. Srivastava, A. Teredesai, R. Thomas, M. Tomko, P. Tziachris, M. Whittaker, H. Wolters, A. Yeo, L. Zhang, and the attendees of the Partnership on AI's April 2019 ABOUT ML workshop for valuable feedback.
Figure. Watch the authors discuss this work in the exclusive Communications video. https://cacm.acm.org/videos/datasheets-for-datasets
2. Bender, E. and Friedman, B. Data Statements for Natural Language Processing: Toward Mitigating System Bias and Enabling Better Science. Trans. of the Assoc. for Computational Linguistics 6 (2018), 587–604.
4. Bolukbasi, T., Chang, K., Zou, J., Saligrama, V., and Kalai, A. Man is to computer programmer as woman is to homemaker? Debiasing Word Embeddings. In Advances in Neural Information Processing Systems (2016).
5. Buolamwini, J. and Gebru, T. Gender shades: Intersectional accuracy disparities in commercial gender classification. In Proceedings of the Conf. on Fairness, Accountability, and Transparency (2018). 77–91.
12. Dastin, J. Amazon scraps secret AI recruiting tool that showed bias against women, 2018; https://reut.rs/3imOH4d.
15. Holstein, K., Vaughan, J., Daumé, H, Dudík, M., and Wallach, H. Improving fairness in machine learning systems: What do industry practitioners need? In Proceedings of 2019 ACM CHI Conf. on Human Factors in Computing Systems.
16. Huang, G., Ramesh, M., Berg, T., and Learned-Miller, E. Labeled Faces in the Wild: A Database for Studying Face Recognition in Unconstrained Environments. Technical Report 07-49. University of Massachusetts Amherst, 2007.
19. Mann, G. and O'Neil, C. Hiring Algorithms Are Not Neutral, 2016; https://hbr.org/2016/12/hiring-algorithms-are-not-neutral.
22. Pang, B. and Lee, L. A sentimental education: Sentiment analysis using subjectivity summarization based on minimum cuts. In Proceedings of the 42nd Annual Meeting of the Assoc. for Computational Linguistics. 2004, 271.
23. Seck, I., Dahmane, K., Duthon, P., and Loosli, G. Baselines and a datasheet for the Cerema AWP dataset. CoRR abs/1806.04016 (2018). http://arxiv.org/abs/1806.04016
25. World Economic Forum Global Future Council on Human Rights 2016–2018. How to Prevent Discriminatory Outcomes in Machine Learning; 2018. https://www.weforum.org/whitepapers/how-to-prevent-discriminatory-outcomes-inmachine-learning.
26. Yagcioglu, S., Erdem, A., Erdem, E., and Ikizler-Cinbis, N. RecipeQA: A challenge dataset for multimodal comprehension of cooking recipes. In Proceedings of the 2018 Conf. on Empirical Methods in Natural Language Processing.
a. We note that in some cases, the people creating a datasheet for a dataset may not be the dataset creators, as was the case with the example datasheets that we created as part of our development process.
Copyright held by authors/owners. Publication rights licensed to ACM.
Request permission to publish from email@example.com
The Digital Library is published by the Association for Computing Machinery. Copyright © 2021 ACM, Inc.
I'm curious if/how the FAIR guiding principles for scientific data management and stewardship align with datasheets for datasets. https://www.nature.com/articles/sdata201618
Also, there seems to be an adversarial relationship between DAIR and industry. Are there opportunities for industry to work with DAIR to help eliminate unwanted societal biases?
This is interesting and timely work. It is crucial that rich context (metadata++) is available for datasets. The 50+ questions are daunting for an author/creator but the more time and effort given to adding this value will accrue on multiple levels: findability; credit for concept development, understandability, and computability/inferencing/linking to advance progress. Libraries have long developed metadata standards for individual objects to serve these value adds, and archives create finding aids that richly contextualize collections. This work illustrates how to move forward with curating datasets. Gary Marchionini
This is a great idea and I wonder if it could be integrated with the Open Science Framework so that people who post their datasets to OSF.io would follow one of your templates. Also, I wondered if this effort could learn from the research reporting guidelines established for medical research (see https://www.equator-network.org/). If you want to submit a medical research article, the journal will demand that you also submit a completed research report like CONSORT (or any of hundreds of others). Maybe data journals could do the same with your template (but not the hundreds part).
Displaying all 3 comments