Pith. sign in

REVIEW 3 major objections 5 minor 21 references

SetGo is a pre-publication metadata-readiness toolkit for scientific AI datasets, claiming to lift FAIR scores from 52–57% to 81–91% across four scientific domains by fixing licensing, provenance, and governance gaps before publication.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 07:31 UTC pith:4NPCXX5I

load-bearing objection Useful pre-publication metadata readiness toolkit, but the headline FAIR gains are self-scored on author-constructed fixtures; the paper candidly admits this, so treat the numbers as illustrative. the 3 major comments →

arxiv 2607.22677 v1 pith:4NPCXX5I submitted 2026-07-10 cs.DL cs.AI

SetGo: Metadata Readiness for Scientific AI Datasets

Sean R. Wilkinson , Polina Shpilker , Wesley Brewer This is my paper
classification cs.DL cs.AI
keywords FAIR datametadata readinessscientific AI datasetsdata governancelicensingprovenancecatalog publicationLLM agents
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

SetGo is a pre-publication metadata-readiness toolkit for scientific AI datasets. The paper establishes that computational readiness and metadata readiness are distinct, and that no single existing tool covers FAIR compliance, licensing, provenance, governance, reproducibility, and catalog readiness before publication. SetGo fills that gap: on fixture records from four large scientific corpora—climate, proteomics, materials, and fusion—guided enrichment lifts overall FAIR scores from 52–57% to 81–91%, and its domain-specific validators expose deficiencies that generic checks miss, such as climate metadata scoring only 4% on an accepted discovery convention. A single publish command produces catalog-ready records and machine-readable metadata sidecars, and an LLM-agent skill lets a scientist run the whole assess-enrich-publish loop in natural language. If the tool behaves as reported, dataset teams could catch metadata gaps before they reach a catalog rather than after.

Core claim

The central claim is that metadata readiness can be measured and repaired before publication with one deterministic toolkit, and that doing so yields large, reproducible FAIR-score gains. SetGo computes a normalized FAIR score as the mean of assessed sub-principles on a 0–3 scale, excluding sub-principles that cannot be evaluated from a local record, then combines that with governance, licensing, provenance, reproducibility, traceability, and catalog-readiness checks. On paired before/after metadata records from four domains, the before state scores 52–57% overall with uniformly low findability (42%), and the after state reaches 81–91%; the proteomics record is capped at 81% because its prot

What carries the argument

The central object is SetGo's six-module assessment-and-repair pipeline, anchored by a transparent FAIR scoring formula: each of the 15 FAIR sub-principles is scored 0–3, the overall score is the equal-weighted mean over the assessed set, and unassessable sub-principles are omitted rather than penalized. A recommender module ranks missing fields by expected score impact, domain validators add standard-specific checks, and catalog exporters turn the enriched metadata record into published records and machine-readable sidecars. The same record flows through license validation, standard provenance ontology construction, reproducibility capture, and traceability links, with an LLM-agent skill fi

Load-bearing premise

The evaluation assumes the paired before/after fixture records faithfully represent the four real datasets, and that SetGo's own FAIR scoring gains predict real-world discoverability and reuse.

What would settle it

Publish the enriched records produced by SetGo for the four corpora and run an independent FAIR-assessment service on those published records; if the independent result falls materially below 81–91%, or if a discovery-and-reuse study finds no measurable improvement, the central claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Dataset teams can run a single metadata-readiness pass before publication and catch gaps that general-purpose FAIR checkers miss, such as near-zero compliance with domain metadata conventions.
  • With guided enrichment, most scientific datasets can reach 81–91% FAIR scores using only a few authoritative values supplied by the scientist, such as a persistent identifier, a license, and a provenance source.
  • A single publish command turns an enriched metadata record into catalog-ready entries and machine-readable metadata sidecars, so the same record can feed multiple registries without re-authoring.
  • LLM-driven agents can execute the assess–enrich–publish loop in natural language, with user confirmation required before authoritative metadata is added, keeping checks deterministic and auditable.
  • Some datasets, such as those derived from protein-structure repositories, may be blocked from full compliance by licensing terms that do not fit standard identifiers, and will need institutional review.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The reported gains are measured by SetGo's own scorer on synthetic fixture records, so the likely next test—which the paper itself flags—is whether external FAIR assessors and real-world discoverability confirm the 81–91% ceiling.
  • If the score gains transfer, metadata readiness could become a standard continuous-integration gate for scientific AI pipelines, blocking publication until a threshold is met, much as code-quality gates work in software development.
  • Because unassessable sub-principles are omitted from the score rather than penalized, the scores may be optimistic; a dataset could appear highly ready while important dimensions such as accessibility via a public registry are simply absent from the local record.
  • A concrete extension would be an independent blinded annotation study comparing before/after records from the tool against expert judgment of publication readiness, to see whether the score deltas correspond to human-perceived improvements.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. SetGo is an open-source Python toolkit for assessing and repairing metadata readiness of scientific AI datasets before publication. The paper positions SetGo as a complement to the REDI computational-readiness pipeline, covering FAIR compliance, licensing, provenance, governance, reproducibility, and catalog readiness in one workflow. SetGo evaluates local metadata.json records, provides domain-aware validators (CF, ACDD, OPTIMADE, PDB/FASTA, IMAS), produces quantitative FAIR scores with letter grades, supports SPDX license validation, PROV-O provenance export, and publishes to Hugging Face Hub, CKAN, or OpenMetadata with Croissant 1.0 sidecars. An LLM-agent skill file (`/setgo`) enables natural-language execution of the assess–enrich–publish loop. The central quantitative claim is that guided enrichment raises overall FAIR scores on four scientific corpora (ERA5, OpenFold, OC22, XGC1) from 52–57% to 81–91% (Abstract, Table 1). The evaluation is fixture-based: the authors construct paired before/after metadata records rather than rerunning pipelines on the original multi-terabyte datasets. The paper explicitly acknowledges in §4.4 that score improvements partially reflect optimization against SetGo's own scoring criteria, that no external FAIR-tool comparison was possible, and that a prospective user study is deferred to future work.

Significance. If the core claims were fully validated, SetGo would fill a genuine gap: no widely available tool appears to combine pre-publication FAIR assessment, licensing/governance checks, provenance generation, and multi-catalog publication in a single open-source package. The system design is sensible, the code and fixtures are open source and reproducible, and the domain-aware validators address real metadata deficiencies that generic FAIR assessors miss (e.g., ACDD compliance for climate data, OPTIMADE species definitions, PDB licensing/SPDX mismatches). These are concrete contributions. However, the headline quantitative result—the 52–57% to 81–91% FAIR score improvement—rests on SetGo's own rule-based scorer applied to author-constructed fixtures, with a scoring rule that omits unevaluable sub-principles from the denominator. The paper is honest about this limitation, but the evaluation as presented does not yet establish that the score gains correspond to externally recognizable improvements in metadata readiness or FAIR compliance. The system contribution is credible and useful; the evaluation is the main weakness and is fixable.

major comments (3)
  1. [§3.1, Eq. (1), Table 1] The headline score gain is computed by SetGo's own scorer, whose definition in Eq. (1) omits sub-principles with score 0 from the mean. For local pre-publication records, remote-resolvability checks such as A2 are systematically excluded from both the before and after scores. This is a lenient convention: an assessor that treats unevaluable FAIR criteria as non-compliant would produce materially lower scores. The paper provides no sensitivity analysis under the alternative convention, so the specific claim that overall FAIR scores rise from 52–57% to 81–91% is not yet established as a robust statement about FAIR compliance. Please report both conventions, or justify the omission rule with evidence that the excluded sub-principles are not relevant to readiness.
  2. [§4 and §4.4] The before/after fixtures are author-constructed. The paper states that after-state fixtures 'reproduce the outputs of actual enrichment sessions,' but no session transcripts, user logs, or independent confirmation are provided. Section 4.4 concedes that the improvements 'partially reflect optimization against SetGo's own scoring criteria.' Consequently, the 81–91% figures could be an artifact of fixture construction and the tool's own heuristics rather than genuine metadata readiness that external tools or users would recognize. The paper already names the missing evidence (external FAIR-tool comparison, expert annotation, downstream discoverability analysis) in §4.4. To support the central claim, at least one external anchor is needed—for example, running F-UJI on published records produced by setgo publish, an independent expert rating of before/after fixtures, or a small user study.
  3. [§2.6 and §4.4] The novelty claim that no prior system covers all six dimensions is plausible but the evaluation does not systematically compare SetGo against any existing tool, even where comparison is possible. The contribution that SetGo 'surfaces deficiencies that general-purpose tools do not detect' is supported only by three illustrative outputs (ACDD 1.3 at 4%, missing OPTIMADE species definition, PDB/SPDX license incompatibility), not by a structured comparison with F-UJI, the FAIR Evaluator, DataLad, or domain-specific validators. Since the authors state that direct comparison is impossible on unpublished local records, the claim should be softened or accompanied by a documented partial comparison on the dimensions that are comparable.
minor comments (5)
  1. [§3, 'ReadinessReport'] The contributions and abstract describe a 'six-dimensional' readiness framework, but the ReadinessReport aggregates 'four scored dimensions' and treats provenance and traceability as structural rather than scored. Please clarify which dimensions are scored and why the two structural dimensions are not included in the composite score.
  2. [§4.1, Table 1] Table 1 reports aggregated percentages but does not show which sub-principles were omitted from the assessed set A for each domain. Adding a per-domain list of omitted sub-principles would improve transparency and help readers interpret the scores.
  3. [Fig. 2 caption] The caption says values are taken from an actual run, but §4.4 notes the absence of logs or transcripts. Consider providing the raw transcript in the repository or clearly labeling Fig. 2 as an illustrative condensation.
  4. [§4.4, Threats to Validity] The sentence 'All scoring rules are grounded in external published standards ... so score improvements reflect genuine metadata additions' overstates the evidence. Grounding in standards is a design property, not a validation result. Suggest rewording to avoid conflating the two.
  5. [References] Reference [1] (WebMCP) is an 'early preview' and may not be a stable citable source. Consider citing a more permanent URL or a versioned document.

Circularity Check

1 steps flagged

Headline FAIR-score gains are self-scored on author-constructed fixtures; no independent measurement chain validates the 52–57% → 81–91% claim.

specific steps
  1. fitted input called prediction [Abstract; §3.1 FAIR Assessment; §4 Evaluation/Table 1; §4.4 Threats to Validity]
    "Guided enrichment raises overall FAIR scores from 52–57% to 81–91% ... we construct paired metadata records representing the state of each dataset before and after guided enrichment ... a sub-principle that cannot be evaluated (score 0) is omitted from the mean, not counted as a zero ... improvements after enrichment partially reflect optimization against SetGo’s own scoring criteria."

    The central quantitative result is not measured by an independent FAIR assessor on real pipeline outputs. The before/after fixtures are author-constructed, the after-state is produced by following SetGo's ranked recommendations (adding DOI, SPDX license, provenance), and the score is computed by SetGo's own rule-based scorer. Adding exactly those fields is what the scorer counts under F1, R1.1, and R1.2, so the score gain is largely imposed by the evaluation design rather than independently validated. The paper concedes this in §4.4. The additional rule that unevaluable sub-principles (e.g., A2) are omitted from the denominator makes the absolute percentages a self-defined measure of assessability, further decoupling the headline numbers from external FAIR compliance benchmarks.

full rationale

The paper's main quantitative claim—that guided enrichment raises FAIR scores from 52–57% to 81–91%—rests on SetGo's own scoring of author-constructed before/after fixtures. The after-state fixtures are generated by following SetGo's recommendations, and the scoring formula counts the added fields, so the improvement is largely wired into the experiment. §4.4 explicitly admits that 'improvements after enrichment partially reflect optimization against SetGo's own scoring criteria.' The scoring rule's exclusion of unevaluable sub-principles from the mean further means the reported percentages are a self-defined measure rather than an independent FAIR audit. This is a substantive circularity in the evaluation of the headline numbers. However, the paper is not merely self-citational: the scoring rules are anchored to external standards (FAIR, SPDX, PROV-O, ACDD, OPTIMADE), the code is open-source, fixtures are included for reproducibility, and qualitative findings such as PDB licensing incompatibility and ACDD 1.3 gaps are concrete checks against published standards. Those aspects give the tool independent content even though the 81–91% figures are not externally validated. No load-bearing self-citation chain was found. Overall, partial circularity: the score gains reduce by construction to the tool's own scoring criteria, but the broader toolkit contributions remain independently meaningful.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The paper's central evaluation rests on self-chosen scoring parameters (equal sub-principle weights, letter-grade thresholds, the exclusion of unassessable sub-principles) and on domain assumptions about the representativeness of fixture records and the validity of FAIR as a readiness proxy. No new physical or theoretical entities are introduced; SetGo is a software artifact. The most consequential assumption is that score increases on SetGo's own metric are meaningful evidence of improved metadata readiness.

free parameters (4)
  • FAIR sub-principle equal weighting = 1/|A|
    The composite score S_FAIR averages assessed sub-principles equally; the authors state equal weighting was chosen by hand to maintain interpretability, not justified by evidence (§4).
  • Letter-grade thresholds and AI-ready cutoff = A≥90, B≥80, C≥70, D≥60, F<60; is_ready at C=70
    Chosen following conventional grading conventions; no empirical basis is provided (§3).
  • Scoring scale and handling of unassessed sub-principles = 0–3 scale; score 0 omitted from mean
    Authors define a four-point scale and exclude unassessable sub-principles (e.g., A2); this can inflate composite scores (§3.1, §4).
  • STRICT policy FAIR threshold = FAIR score ≥ 0.7
    Configurable policy level; threshold chosen ad hoc for the governance module (§3.2).
axioms (4)
  • domain assumption FAIR principles and their 15 sub-principles are a valid and sufficient measure of metadata readiness for publication.
    The entire scoring framework treats FAIR compliance as the core readiness metric; §3.1 maps every sub-principle to a 0–3 score.
  • domain assumption The REDI metadata.json schema is canonical; other pipelines can be adapted with lightweight adapters.
    SetGo treats metadata.json as primary input; §4.4 limitations state other pipelines require adapters.
  • domain assumption The constructed before/after fixture records faithfully represent the real metadata of the four benchmark corpora.
    §4 states before-state fixtures are 'drawn directly from REDI runs' and after-state fixtures 'reproduce outputs of actual enrichment sessions,' but no evidence is given that these fixtures capture the variability of the 4.3–106 TB corpora.
  • domain assumption User-supplied identifiers, licenses, and provenance links are correct and authoritative.
    Score improvements assume user-supplied values are valid; no verification against external registries beyond SPDX syntax is described (§4.3).

pith-pipeline@v1.3.0-alltime-deepseek · 9467 in / 13587 out tokens · 142563 ms · 2026-08-02T07:31:37.634746+00:00 · methodology

0 comments
read the original abstract

Scientific datasets intended for AI use require both computational readiness for model training and metadata readiness for discovery, sharing, and reuse. The Readiness Engine for Data Integration (REDI) addresses computational readiness, but no corresponding tool evaluates whether a dataset's metadata are sufficiently complete, governed, and standards-compliant for publication and agent-based consumption. Existing FAIR assessors operate only on published repository records, and no single system covers FAIR compliance, licensing, provenance, governance, reproducibility, and catalog readiness together. We present SetGo, an open-source Python toolkit that assesses and repairs metadata readiness across these six dimensions before a dataset is published or archived. Applied to four scientific corpora, SetGo surfaces deficiencies that general-purpose tools do not detect: ERA5 climate metadata scores 4% on ACDD 1.3 compliance; materials datasets fail OPTIMADE species-definition requirements; and PDB-derived proteomics data carries licensing terms incompatible with standard SPDX identifiers. Guided enrichment raises overall FAIR scores from 52-57% to 81-91%, and a single setgo publish command pushes to Hugging Face Hub, CKAN, or OpenMetadata with ML Commons Croissant 1.0 metadata sidecars. To support interactive and automated workflows, SetGo integrates with coding agents powered by large language models (LLMs) through a /setgo skill that enables natural-language execution of the full assess-enrich-publish loop, with user involvement limited to supplying missing metadata values.

Figures

Figures reproduced from arXiv: 2607.22677 by Polina Shpilker, Sean R. Wilkinson, Wesley Brewer.

Figure 1
Figure 1. Figure 1: The REDI–SetGo iterative data readiness lifecycle [ [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Left: Raw setgo fair output for an ERA5 climate dataset, showing a 57% initial FAIR score and three NOT_COMPLIANT principles. Right: An LLM agent guided by SKILL.md executes the full assess–enrich–publish loop from a single natural-language request, calling setgo fair --style json and setgo publish internally while prompting the scientist only for the three missing metadata fields (transcript condensed for… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

21 extracted references · 3 canonical work pages · 2 internal anchors

  1. [1]

    WebMCP is available for early preview

    2026. WebMCP is available for early preview. https://developer.chrome.com/ blog/webmcp-epp. Accessed: 2026-05-07

    2026

  2. [2]

    Gustaf Ahdritz, Nazim Bouatta, Christina Floristean, Sachin Kadyan, Qinghui Xia, et al. 2024. OpenFold: Retraining AlphaFold2 Yields New Insights into Its Learning Mechanisms and Capacity for Generalization.Nature Methods21 (2024), 1514–1524. doi:10.1038/s41592-024-02272-z

    doi:10.1038/s41592-024-02272-z 2024

  3. [3]

    Anthropic. 2025. Claude Code. https://github.com/anthropics/claude-code. Ac- cessed: 2026-05-06

    2025

  4. [4]

    Wesley Brewer, Patrick Widener, Valentine Anantharaj, Feiyi Wang, Tom Beck, Arjun Shankar, and Sarp Oral. 2026. Data Readiness Pipeline Patterns for Scientific AI at Scale: Insights from Climate, Fusion, Life Sciences, and Materials.AI Magazine47, 1 (2026). doi:10.1002/aaai.70056

    doi:10.1002/aaai.70056 2026

  5. [5]

    Parker, Sadnan Al Manir, Edilberto Amorim, James Eddy, et al

    Timothy Clark, Harry Caufield, Jillian A. Parker, Sadnan Al Manir, Edilberto Amorim, James Eddy, et al. 2024. AI-Readiness for Biomedical Data: Bridge2AI Recommendations.bioRxiv(2024). doi:10.1101/2024.10.23.619844

    doi:10.1101/2024.10.23.619844 2024

  6. [6]

    Anusuriya Devaraju and Robert Huber. 2021. An Automated Solution for Measuring the Progress toward FAIR Research Data.Patterns2, 11 (2021). doi:10.1016/j.patter.2021.100370

    arXiv 2021

  7. [7]

    Wilkinson, Finn Bacall, Stian Soiland- Reyes, Simone Leo, Luca Pireddu, Stuart Owen, Nick Juty, JoséM

    Ove Johan Ragnar Gustafsson, Sean R. Wilkinson, Finn Bacall, Stian Soiland- Reyes, Simone Leo, Luca Pireddu, Stuart Owen, Nick Juty, JoséM. Fernández, Tom Brown, Hervé Ménager, Björn Grüning, Salvador Capella-Gutierrez, Frederik Coppens, and Carole Goble. 2025. WorkflowHub: a registry for computational workflows.Scientific Data12, 1 (May 2025). doi:10.103...

    doi:10.1038/s41597-025-04786-3 2025

  8. [8]

    Halchenko, Kyle Meyer, Benjamin Poldrack, Debanjum Singh Solanky, Adina S

    Yaroslav O. Halchenko, Kyle Meyer, Benjamin Poldrack, Debanjum Singh Solanky, Adina S. Wagner, Jason Gors, Dave MacFarlane, Dorian Pustina, Vanessa Sochat, Satrajit S. Ghosh, et al. 2021. DataLad: distributed system for joint management of code, data, and their relationship.Journal of Open Source Software6, 63 (2021),

    2021

  9. [9]

    Kaveen Hiniduma, Suren Byna, and Jean Luca Bez. 2025. Data Readiness for AI: A 360-Degree Survey.Comput. Surveys(2025). doi:10.1145/3722214

    doi:10.1145/3722214 2025

  10. [10]

    Kaveen Hiniduma, Suren Byna, Jean Luca Bez, and Ravi Madduri. 2024. AI Data Readiness Inspector (AIDRIN) for Quantitative Assessment of Data Readiness for AI. InProceedings of the 36th International Conference on Scientific and Statistical Database Management. doi:10.1145/3676288.3676296

    arXiv 2024

  11. [11]

    Data Readiness Levels

    Neil D. Lawrence. 2017. Data Readiness Levels.arXiv preprint arXiv:1705.02245 (2017). doi:10.48550/arXiv.1705.02245

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1705.02245 2017

  12. [12]

    Massimiliano Lupo Pasini, Jong Youl Choi, Kshitij Mehta, Pei Zhang, David Rogers, et al. 2025. Scalable Training of Trustworthy and Energy-Efficient Predic- tive Graph Foundation Models for Atomistic Materials Modeling: A Case Study with HydraGNN.The Journal of Supercomputing(2025). doi:10.1007/s11227-025- 07029-9

    doi:10.1007/s11227-025- 2025

  13. [13]

    Gupta, and Aditya Grover

    Tung Nguyen, Johannes Brandstetter, Ashish Kapoor, Jayesh K. Gupta, and Aditya Grover. 2023. ClimaX: A Foundation Model for Weather and Climate.arXiv preprint arXiv:2301.10343(2023). doi:10.48550/arXiv.2301.10343

    Pith/arXiv arXiv doi:10.48550/arxiv.2301.10343 2023

  14. [14]

    Renan Souza, Tyler J Skluzacek, Sean R Wilkinson, Maxim Ziatdinov, and Rafael Ferreira da Silva. 2023. Towards Lightweight Data Integration using Multi-workflow Provenance and Data Observability. InIEEE International Con- ference on e-Science. doi:10.1109/e-Science58273.2023.10254822

    arXiv 2023

  15. [15]

    United States Department of Energy. 2025. Genesis Mission. https://genesis. energy.gov/. Accessed: 2026-06-18

    2025

  16. [16]

    Wilkinson, Michel Dumontier, IJsbrand Jan Aalbersberg, Gabrielle Apple- ton, Myles Axton, Arie Baak, Niklas Blomberg, Jan-Willem Boiten, Luiz Bonino da Silva Santos, Philip E

    Mark D. Wilkinson, Michel Dumontier, IJsbrand Jan Aalbersberg, Gabrielle Apple- ton, Myles Axton, Arie Baak, Niklas Blomberg, Jan-Willem Boiten, Luiz Bonino da Silva Santos, Philip E. Bourne, et al. 2016. The FAIR Guiding Principles for scientific data management and stewardship.Scientific Data3 (2016), 160018. doi:10.1038/sdata.2016.18

    doi:10.1038/sdata.2016.18 2016

  17. [17]

    Wilkinson, Meznah Aloqalaa, Khalid Belhajjame, Michael R

    Sean R. Wilkinson, Meznah Aloqalaa, Khalid Belhajjame, Michael R. Crusoe, Bruno de Paula Kinoshita, Luiz Gadelha, Daniel Garijo, Ove Johan Ragnar Gustafs- son, Nick Juty, Sehrish Kanwal, Farah Zaib Khan, Johannes Köster, Karsten Peters- von Gehlen, Line Pouchard, Randy K. Rannow, Stian Soiland-Reyes, Nicola So- ranzo, Shoaib Sufi, Ziheng Sun, Baiba Vilne,...

    2025

  18. [18]

    Automated Data Readiness for Scientific AI

    Sean R. Wilkinson, Valentine G. Anantharaj, Jong Youl Choi, Ketan Maheshwari, Marshall McDonnell, Massimiliano Lupo Pasini, Polina Shpilker, Renan Souza, Patrick Widener, Sarp Oral, and Wesley Brewer. 2026. Automated Data Readiness for Scientific AI.arXiv preprint arXiv:2607.02771(2026). doi:10.48550/arXiv.2607. 02771

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2607.02771 2026

  19. [19]

    Wilkinson, Greg Eisenhauer, Anuj J

    Sean R. Wilkinson, Greg Eisenhauer, Anuj J. Kapadia, Kathryn Knight, Jeremy Logan, Patrick Widener, and Matthew Wolf. 2022. F*** workflows: when parts of FAIR are missing. In2022 IEEE 18th International Conference on e-Science. 507–512. doi:10.1109/eScience55777.2022.00090

    arXiv 2022

  20. [20]

    Matei Zaharia, Andrew Chen, Aaron Davidson, Ali Ghodsi, Sue Ann Hong, Andy Konwinski, Siddharth Murching, Tomas Nykodym, Paul Ogilvie, Mani Parkhe, et al. 2018. Accelerating the Machine Learning Lifecycle with MLflow.IEEE Data Engineering Bulletin41, 4 (2018), 39–45

    2018

  21. [3262]

    doi:10.21105/joss.03262

    work page doi:10.21105/joss.03262