PaperlayerView exact revision on arXiv

Source-linked AI summary

Retiring Adult: New Datasets for Fair Machine Learning

Frances Ding, Moritz Hardt, John Miller, Ludwig Schmidt

arXiv:2108.04884v3cs.LGstat.ML

TL;DR

Fairness research on tabular data relies heavily on UCI Adult, whose idiosyncratic income threshold may limit its external validity. This paper builds new Census-derived prediction datasets and finds that fairness findings vary with the income threshold.

  • Problem

    Fairness research on tabular data relies heavily on UCI Adult, whose idiosyncratic income threshold may limit its external validity.

    Source: §1 · p. 1, §2.2 · p. 4

  • Method

    The authors create a suite of US Census-derived datasets for prediction tasks across several domains.

    Source: §1 · p. 2

  • Results

    Fairness findings are sensitive to income-threshold choice: fairness-criteria violations, trade-offs, and intervention effectiveness vary significantly with the threshold.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Fairness interventions should specify their locus, such as national or state-by-state modeling, because results differ significantly.

    Source: §1 · p. 3

    The paper frames its contribution as benchmark data, distinguishing benchmarks from substantive domain-specific investigations.

    Source: §3.2 · p. 8

Abstract

from arXiv · show

Although the fairness community has recognized the importance of data, researchers in the area primarily rely on UCI Adult when it comes to tabular data. Derived from a 1994 US Census survey, this dataset has appeared in hundreds of research papers where it served as the basis for the development and comparison of many algorithmic fairness interventions. We reconstruct a superset of the UCI Adult data from available US Census sources and reveal idiosyncrasies of the UCI Adult dataset that limit its external validity. Our primary contribution is a suite of new datasets derived from US Census surveys that extend the existing data ecosystem for research on fair machine learning. We create prediction tasks relating to income, employment, health, transportation, and housing. The data span multiple years and all states of the United States, allowing researchers to study temporal shift and geographic variation. We highlight a broad initial sweep of new empirical insights relating to trade-offs between fairness criteria, performance of algorithmic interventions, and the role of distribution shift based on our new datasets. Our findings inform ongoing debates, challenge some existing narratives, and point to future research directions. Our datasets are available at https://github.com/zykls/folktables.

1 Introduction

The paper reconstructs UCI Adult and introduces Census-derived datasets for fair-ML research, using them to show that fairness findings depend on context, intervention choices, and the data examined.

Introduction

  • UCI Adult dominates fairness research on tabular data, and reconstructing a superset from Census records revealed a prediction-task idiosyncrasy that limits its external validity.The dataset has appeared in more than 300 fairness-related papers and has supported the development and comparison of many interventions.Source: §1 · pp. 1–2
  • UCI Adult’s $50k income threshold is the 76th US individual-income quantile, the 88th for Black people, and the 89th for women, and fairness findings vary with threshold choice.The authors report that this extreme threshold can understate and misrepresent the broader picture of fairness violations, trade-offs, and intervention effectiveness.Source: §1 · p. 2
  • The new Census-derived datasets cover income, employment, health, transportation, and housing across multiple years and all US states, enabling study of temporal shift and geographic variation.The authors also release folktables, a Python package for accessing the prediction tasks and creating new ones through a simple API.Source: §1 · p. 2
  • Fairness criteria and intervention effects vary substantially by state, cross-state accuracy and fairness changes are unpredictable, and fairness criteria are more stable over time than predictive accuracy.These findings indicate that claims about fairness require careful contextual qualification and that tools developed in one setting may not transfer gracefully to another.Source: §1 · p. 2
  • Fairness interventions must specify whether the model is trained on the full US population or state by state, because results differ significantly and guidance on this choice remains lacking.Source: §1 · p. 3
  • Increasing dataset size does not necessarily reduce observed disparities, nor does social progress as measured by years passed.Source: §1 · p. 3

2 Archaeology of UCI Adult: Origin, Impact, Limitations

UCI Adult became a widely used fairness benchmark, but its unclear provenance and idiosyncratic income threshold limit what its results can establish. The authors reconstruct it from 1994 CPS data and show that fairness and accuracy results vary with the threshold choice.

2 Archaeology of UCI Adult: Origin, Impact, Limitations

  • Sensitive attributes helped establish UCI Adult’s role in fairness research: Calders et al. used it in 2009, and Zemel et al. popularized fair representations as a standard baseline.Pedreschi et al. earlier used the smaller UCI German credit dataset, which became less widely used.Source: §2 · p. 3
  • UCI Adult became a standard fairness test case for criteria, interventions, methods, and causal modeling, appearing in more than 300 fairness papers and examples in AI Fairness 360 and Fairlearn.It is the UCI repository’s second most popular dataset among more than 500 datasets.Source: §2 · pp. 3–4

2.1 Reconstruction of UCI Adult

  • The authors reconstructed IPUMS Adult from the 1994 CPS ASEC, matching variables through documented encoding maps and nearest-neighbor row matching, with exact matches for 13 of 15 variables.The reconstruction contains 49,531 rows versus UCI Adult’s 48,842; the difference likely reflects UCI’s use of fnlwgt in inclusion criteria, which could not be exactly matched.Source: §2.1 · p. 4

2.2 Varying income threshold

  • The $50,000 income cutoff is idiosyncratic: it is the 76th income quantile overall and the 88th and 89th for Black and female populations, yielding unusually high subgroup accuracy.Logistic regression accuracy is 85% overall, 91.4% on Black instances, and 92.7% on female instances; almost all Black and female instances fall below the threshold.Source: §2.2 · pp. 4–5
  • Accuracy and fairness-violation magnitudes vary substantially with the income threshold; intervention trade-offs also change, and high thresholds widen equal-opportunity confidence intervals.Near $25,000, the accuracy drop required to achieve demographic parity or equal opportunity is significantly larger than near $50,000; for in-processing demographic parity, equal-opportunity violation increases monotonically as the threshold rises.Source: §2.2 · pp. 5–6, Figure 1 · p. 5

3 New datasets for algorithmic fairness

The authors introduce ACS PUMS-based prediction tasks to broaden empirical evaluation in algorithmic fairness beyond UCI Adult. The tasks support geographic and temporal comparisons, while the paper frames them as benchmarks for studying algorithms rather than as substantive investigations of inequality.

3 New datasets for algorithmic fairness

  • The US Census Bureau’s high-quality collection and legally mandated surveys provide a strong data foundation for new fairness-evaluation tasks.The authors use the Census data ecosystem to support stronger empirical evaluation practices in algorithmic fairness.Source: §3 · p. 6

3.1 Available prediction tasks

  • ACS PUMS yields five tasks: ACSIncome, ACSPublicCoverage, ACSMobility, ACSEmployment, and ACSTravelTime.They predict income above $50,000; public insurance coverage; residential-address stability; employment; and commutes longer than 20 minutes. The tasks filter populations by age, income, work, or other criteria, including low-income people under 65 for public coverage and adults ages 18–35 for mobility.Source: §3.1 · p. 7
  • Each task includes age, race, and sex features and can be instantiated across 50 states and Puerto Rico and five years, yielding 255 datasets per task.The authors also provide US-wide datasets by concatenating state data, enabling geographic and temporal comparisons of distribution shift. The tasks range from problems with higher accuracy than a constant predictor to the lower-signal ACSMobility task.Source: §3.1 · p. 7

3.2 Scope and limitations

  • The datasets are intended as benchmarks for empirical study of algorithmic fairness, not as investigations of substantive inequality in domains such as income or employment.The authors envision evaluations of methods, fairness criteria, causal models, and transfer across contexts, complementing theoretical work and reducing reliance on UCI Adult.Source: §3.2 · pp. 7–8
  • A key limitation is that the new datasets are entirely US-centric, leaving international contexts underrepresented in the fairness data ecosystem.The authors specifically note the need for stronger engagement with empirical work in the Global South [1].Source: §3.2 · p. 8

4 A tour of empirical observations

The authors present a non-exhaustive initial sweep using ACS PUMS tasks to study population variation, intervention locus, and whether larger or later datasets mitigate disparities. The experiments illustrate how these datasets can support broader empirical evaluation of fairness questions.

4.1 Variation within the population

  • On ACSIncome demographic-parity interventions, effects vary markedly by state: LFR reduces violation by 10% in some states and close to zero in others, while post-processing accuracy drops range from less than 1% to closer to 5%.Source: §4.1 · p. 10
  • Geographic transfer is unpredictable: accuracy and demographic-parity violations vary across test states, and post-processing that achieves parity in one state may fail elsewhere.Figure 3 compares an unconstrained GBM and a post-processed GBM trained in one state and tested in-distribution and across the other 49 states.Source: Figure 3 · p. 9, §4.1 · p. 10
  • Fairness criteria remain essentially constant over 2014–2018 even as accuracy declines slightly, including for models using LFR, ExpGrad, or post-processing.Models were trained on US-wide 2014 ACSIncome data; those satisfying demographic parity in 2014 continued to satisfy it on 2015–2018 data.Source: §4.1 · p. 10, Figure 4 · p. 11

4.2 Specifying a locus of intervention

  • Applying LFR or post-processing state by state improves US-wide accuracy versus US-level application while preserving demographic parity with post-processing or further mitigating its violations with LFR.The comparison uses a GBM base classifier and aggregates state-level results across the US.Source: §4.2 · p. 10, Table 2 · p. 11

4.3 Increased dataset size doesn’t necessarily mitigate observed disparities

  • Larger, newer, and more socially progressive datasets do not automatically mitigate disparities: a GBM trained on ACSIncome rather than IPUMS Adult does not improve true-positive-rate disparities.The authors attribute persistent disparities to social inequality and nonzero Bayes error, arguing that dataset and model design must directly address how classification errors are distributed.Source: §4.3 · p. 12, Table 3 · p. 12

5 Discussion and future directions · A Adult reconstruction · A.1 Additional reconstruction details

The authors present their empirical observations as prompts for further evaluation of fairness methods, distribution shift, and causal inference. Their Adult reconstruction required exact matches in IPUMS CPS, with specific exceptions for “native-country” and “fnlwgt.”

5 Discussion and future directions

  • The authors call for broad evaluation of existing methods and conjecture that none will outperform threshold adjustment for demographic or error rate parity on their datasets.They suggest a well-tuned tree ensemble may be close to the Bayes error bound, making threshold adjustment optimal under Theorem 5.3 in [20]; how close their tasks are to Bayes optimal remains an open question.Source: §5 · p. 12
  • They identify distribution shift as a research priority, asking about consistency across geographies, greater temporal than geographic stability, and transfer to deployment contexts relevant to policy.They also ask what sensitivity to shift implies for algorithmic tools developed in one context and deployed in another.Source: §5 · p. 12
  • The datasets could also test causal inference methods, including how invariant risk minimization [3] performs across geographic contexts.The authors note that they have not yet explored this direction.Source: §5 · pp. 12–13

A.1 Additional reconstruction details

  • The reconstruction included a candidate variable only when an exact nearest-neighbor match could be found in IPUMS CPS for each UCI Adult row matching prior variables.The authors report two exceptions to this rule.Source: §A.1 · p. 16
  • For “native-country,” Russia and Guyana codes in IPUMS CPS were mapped to “unknown” to match all UCI Adult rows, though the discrepancy is unexplained.The UCI Adult documentation does not list Russia or Guyana as possible values.Source: §A.1 · p. 16
  • The authors found no exact IPUMS CPS match for “fnlwgt,” a Census Bureau sample weight rather than an individual demographic feature, and excluded it from experiments.UH_WGTS_A1 was the closest match and had a similar distribution.Source: §A.1 · pp. 16–17

A.2 Varying the income threshold experiments

The experiments vary the income threshold from $6,000 to $72,000 using a reconstructed IPUMS Adult train/test split. Under equality of opportunity, accuracy costs and fairness trade-offs vary substantially with the threshold, while high thresholds make intervention comparisons less certain.

A.2 Varying the income threshold experiments

  • The study varies the income-prediction threshold from $6,000 to $72,000, using 32,094 training and 13,755 test examples from the 49,531-example reconstruction.For each threshold, the task is to predict whether income exceeds that amount; categorical features are one-hot encoded, Education-Num and Age are clustered, and features are standardized.Source: §A.2 · p. 17
  • The unconstrained baseline is gradient boosting, while the compared fairness interventions are LFR, reductions, and post-processing, with LFR omitted for equality of opportunity.The baseline uses exponential loss with five estimators and max_depth 5; LFR parameters were selected by grid search at threshold 50,000, and the reductions and post-processing implementations use their specified defaults or provided implementations.Source: §A.2 · p. 17
  • As the threshold varies, the accuracy drop needed to enforce equality of opportunity and the trade-off between criteria vary substantially; high thresholds widen confidence intervals and hinder intervention comparisons.The results are described as very similar to the demographic-parity experiment; LFR is excluded because it does not enforce equality of opportunity without modification.Source: §A.2 · p. 17

B New prediction task details … B.5 ACSTravelTime

This appendix specifies prediction-task definitions, including targets, features, encodings, and population filters, for tasks spanning income, public health coverage, mobility, employment, and travel time.

B New prediction task details

  • The task descriptions list each feature’s ACS PUMS variable code, extended description, and value range.More feature information is available in the ACS PUMS documentation.Source: B New prediction task details · p. 17

B.1 ACSIncome

  • ACSIncome predicts whether working adults’ yearly income exceeds $50,000, labeling PINCP > 50000 as 1; the threshold can be changed.The task filters for age over 16, income over 100, usual work hours over 0, and person weight at least 1.Source: §B.1 · pp. 18, 21

B.2 ACSPublicCoverage

  • ACSPublicCoverage predicts whether a low-income person not eligible for Medicare has public health coverage, labeling PUBCOV == 1 as positive.The appendix describes each task’s features and filters and points readers to ACS PUMS documentation for feature details.Source: §B.2 · p. 21
  • The listed criteria require age (AGEP) to be less than 65 and total person’s income (PINCP) to be less than $30,000.Source: §B.2 · p. 25

B.3 ACSMobility

  • ACSMobility predicts whether a young adult moved addresses in the last year, labeling MIG == 1 as positive and all other values as negative.The appendix lists coded ranges for the task’s demographic and household variables.Source: §B.3 · p. 25
  • ACSMobility records usual weekly hours as 1–98 or 99 for 99 or more, with N/A for people under 16 or who did not work in the past year.Source: §B.3 · p. 30
  • JWMNP (Travel time to work) is N/A for nonworkers or workers who worked at home; otherwise, values are integer minutes from 1 to 200, with values above 200 top-coded at 200.Source: §B.3 · p. 30
  • The ACSMobility age criterion is greater than 18 and less than 35.Source: §B.3 · p. 30

B.4 ACSEmployment

  • ACSEmployment predicts adult employment, labeling a person 1 when ESR == 1 and 0 otherwise.Source: §B.4 · p. 31
  • Ancestry recode value 8 is suppressed for data year 2018 in select PUMAs.Source: §B.4 · p. 33
  • The ACSEmployment data require age above 16 and person weight of at least 1.Source: §B.4 · p. 35

B.5 ACSTravelTime

  • ACSTravelTime labels working adults as positive when JWMNP exceeds 20 minutes, and negative otherwise.The task is restricted to people older than 16, with person weight at least 1 and employment status coded as employed.Source: §B.5 · pp. 35, 39

B.6 Dataset access and license

The authors provide software for downloading ACS PUMS data and constructing the paper’s prediction tasks or new tasks from ACS PUMS products. The data and IPUMS reconstruction are governed by their respective terms of use.

B.6 Dataset access and license

  • The software package downloads ACS PUMS data and constructs the paper’s prediction tasks and new tasks using ACS PUMS products.Source: §B.6 · p. 39
  • ACS PUMS data are subject to US Census Bureau terms of service, while the IPUMS adult reconstruction is subject to IPUMS terms of use.Source: §B.6 · p. 39

B.7 Table 1 experiment details

The Table 1 experiments use 2018 US-Wide ACS PUMS data with state-level sampling, an 80/20 train-test split, standardized features, and three classifier baselines.

B.7 Table 1 experiment details

  • Experiments for five ACS tasks use 2018 US-Wide ACS PUMS data, capped at 100,000 examples per state with random subsampling for larger states.Source: §B.7 · p. 39
  • The data are randomly split into 80% training and 20% test examples, and all features are standardized to zero mean and unit variance.Source: §B.7 · p. 39
  • The baselines are a majority-class Constant Predictor, logistic regression, and gradient-boosted decision trees, implemented with Pedregosa et al.’s [28] defaults.Source: §B.7 · p. 39

C Tour of empirical observations: missing experimental details

The empirical observations use a shared GBM base model and ACSIncome data, with state-based train/test splits except in experiments on temporal distribution shift. The paper reports confidence intervals for accuracy and fairness violations.

Models and hyperparameters

  • All experiments use GBM, selected for fast training and higher unconstrained accuracy than logistic regression and random forests; other base models gave qualitatively similar results.The GBM uses exponential loss, 5 estimators, and max_depth 5, with other hyperparameters default; these settings were selected by a small accuracy grid search on ACSIncome. The authors use LFR, ExpGrad [2], and Hardt et al.’s [20] threshold adjustment for demographic parity, with equality-of-opportunity experiments in Appendix D.Source: C Tour of empirical observations:… · pp. 39–40

Datasets

  • Experiments use ACSIncome and 2018 1-Year ACS PUMS data, randomly splitting each state 80% for training and 20% for testing, except temporal-shift experiments use 2014–2017 data.The US-Wide dataset combines the state training and testing sets over all 50 states and Puerto Rico.Source: C Tour of empirical observations:… · p. 40

Confidence intervals

  • Accuracy is reported with 95% Clopper–Pearson intervals, while demographic-parity and equality-of-opportunity violations use 95% Newcombe intervals for differences between two binomial proportions.These intervals account for random variation in estimating the metrics.Source: C Tour of empirical observations:… · p. 40

D Additional experiments · D.1 Intervention effect sizes across states · D.2 Geographic distribution shift

The appendix repeats Section 4’s experiments on five additional prediction tasks, keeping the experimental details unchanged. It examines how intervention effects vary across states and extends geographic-shift experiments to additional ACS tasks.

D Additional experiments

  • The appendix’s main extension repeats Section 4’s experiments on five other prediction tasks using the same models, hyperparameters, and other experimental details.Source: D Additional experiments · p. 40

D.1 Intervention effect sizes across states

  • For each state, the ACSIncome experiment compares an unconstrained GBM with versions using LFR, fair reductions [2], or group-based threshold adjustment [20] for equality of opportunity.Source: §D.1 · p. 40
  • Figure 6 shows that fairness-intervention effect sizes vary by state for ACSIncome, with each arrow comparing accuracy and equality-of-opportunity violation before and after intervention.Source: Figure 6 · p. 41
  • The demographic-parity intervention experiment is also conducted on ACSPublicCoverage, ACSEmployment, ACSMobility, and ACSTravelTime.Source: §D.1 · p. 41

D.2 Geographic distribution shift

  • For ACSIncome, the geographic-shift experiment compares unconstrained and post-processed GBMs on same-state test data and test sets from the other 49 states, using equality of opportunity with respect to race.Source: §D.2 · p. 41
  • The corresponding demographic-parity geographic-shift experiments cover ACSPublicCoverage, ACSEmployment, ACSMobility, and ACSTravelTime.Source: §D.2 · p. 41

D.3 Temporal distribution shift · E Datasheet · E.1 Motivation

Fairness violations remain essentially constant over time even as accuracy drops in most problems, while state-to-state transfer is unpredictable. The datasheet documents the Census-based prediction tasks created to extend the ecosystem for algorithmic fairness research.

D.3 Temporal distribution shift

  • State-to-state transfer produces unpredictable accuracy and fairness outcomes for ACSIncome and four other ACS prediction tasks, both with unconstrained models and equality-of-opportunity post-processing.The figures compare models trained in one state with in-distribution evaluation there and out-of-distribution evaluation in the other 49 states.Source: Figure 8 · p. 43, Figure 9 · p. 44, Figure 10 · p. 45
  • Accuracy drops over time for most problems, while equality-of-opportunity and demographic-parity violations remain essentially constant across the evaluated tasks.The experiments train on US-wide 2014 data; equality of opportunity is tested for ACSIncome, and demographic parity for ACSPublicCoverage, ACSMobility, ACSEmployment, and ACSTravelTime.Source: §D.3 · p. 41, Figure 13 · p. 47, Figure 14 · p. 48

E Datasheet

  • The datasheet covers both the introduced prediction tasks and their underlying US Census data sources, often referring readers to existing Census documentation.Source: E Datasheet · p. 49
  • The datasheet points readers to the Folktables repository for its most up-to-date version.Source: E Datasheet · p. 49

E.1 Motivation

  • The new US Census-based prediction tasks were created to extend the dataset ecosystem available for algorithmic fairness research.Source: §E.1 · p. 49
  • The datasheet identifies Frances Ding, Moritz Hardt, John Miller, and Ludwig Schmidt as creators; the first three were employed by UC and Ludwig Schmidt by Toyota Research during the project.The research was funded by grants administered through the University of California.Source: §E.1 · p. 49

E.2 Composition · E.3 Collection process · E.4 Preprocessing / cleaning / labeling

The datasets comprise person-level samples from US Census surveys, with survey-specific features and labels. The paper documents how those surveys were collected and how the released prediction tasks are constructed and preprocessed.

E.2 Composition

  • The reconstructed IPUMS Adult and ACS datasets represent individuals; Census source data also include household information and household–individual relationships.The released datasets themselves contain no relationships between individuals.Source: §E.2 · pp. 49–50
  • IPUMS Adult contains 49,531 rows, and both it and the ACS datasets are samples of the US population whose per-instance weights matter for population representation.The paper refers readers to Census documentation on representativeness.Source: §E.2 · p. 49
  • IPUMS Adult uses the original UCI Adult features and an income label with continuous values available; ACS instances use features and labels documented in Appendix B.Some features, including country of origin in IPUMS Adult, have missing values; individuals have attributes such as age, gender, and race.Source: §E.2 · pp. 50–51

E.3 Collection process

  • Census survey responses supply the data: ACS uses internet, mail, telephone, and in-person interviews, while CPS uses telephone and in-person interviews; both collect annually.ACS samples housing units uniformly within each county, while CPS samples housing units from specified sampling areas.Source: §E.3 · p. 52
  • ACS participation is mandatory, whereas CPS participation is voluntary with consent obtained at interview; the authors know of no Census Bureau mechanism to revoke consent.The Census Bureau assesses privacy risks and uses statistical disclosure control; the authors state that their derived tasks do not increase privacy risks.Source: §E.3 · p. 53

E.4 Preprocessing / cleaning / labeling

  • The IPUMS Adult reconstruction uses CPS ASEC data without altering selected variables; Census releases top-code variables and impute some missing values in both CPS and ACS.ACS tasks further process data, usually mapping NaNs to −1, and define binary labels by discretizing target variables; task preprocessing and labeling can be changed.Source: §E.4 · p. 54
  • The package provides access to Census-released data and code to define prediction tasks, but raw survey answers are not publicly released due to privacy considerations.The paper notes that the prediction-task code is available through the folktables project.Source: §E.4 · p. 54

E.5 Uses · E.6 Distribution · E.7 Maintenance

The sections describe five ACS-based prediction tasks and their intended limits, then specify how the datasets are distributed and maintained. They warn that small state-level subgroup samples can undermine conclusions and fairness benchmarking.

E.5 Uses

  • The paper creates five ACS PUMS tasks: ACSIncome, ACSPublicCoverage, ACSMobility, ACSEmployment, and ACSTravelTime.They predict, respectively, income above $50,000, public coverage for low-income people not eligible for Medicare, moving in the past year, adult employment, and working-adult travel time over 20 minutes.Source: §E.5 · pp. 54–55
  • Users can define other ACS PUMS tasks with different features or targets to benchmark models across varied Bayes error rates and subgroup base-rate disparities.Source: §E.5 · p. 55
  • Small state-level subgroup samples can make generalizations inaccurate and fairness benchmarks superficial; users should not re-identify people or treat these datasets as substantive evidence on inequality or demographic shifts.The datasets are intended primarily for benchmarking machine-learning algorithms.Source: §E.5 · pp. 55–56

E.6 Distribution

  • The datasets are distributed publicly through GitHub, with release planned for August 1, 2021; they have no DOI.Source: §E.6 · p. 56
  • The folktables package and loading code use the MIT license, while the ACS data and IPUMS Adult reconstruction follow their respective third-party terms of use.The paper states that no export controls or regulatory restrictions are known to apply.Source: §E.6 · pp. 56–57

E.7 Maintenance

  • The folktables team hosts and maintains the dataset on GitHub, updates it as needed based on errors and community feedback, and posts errata there.Source: §E.7 · p. 57
  • The ACS-based folktables data and CPS-based Adult reconstruction inherit and will respect their source datasets’ retention policies.Source: §E.7 · p. 57
  • Older versions will remain supported on GitHub with version metadata and releases; contributions are reviewed by maintainers and distributed in new releases.Source: §E.7 · p. 58

Related papers

Loading 2108.04884v3…