Every statistical threshold, noise-estimation formula, and metadata schema referenced on bestaidata.org traces to peer-reviewed computer science literature, federal risk frameworks, and open-source specification bodies.
Timnit Gebru, Jamie Morgenstern, Briana Vecchione, Jennifer Wortman Vaughan, Hanna Wallach, Hal Daume III, Kate Crawford (Communications of the ACM / arXiv). Foundational specification establishing standardized dataset lifecycle documentation across collection motivation, composition, consent, preprocessing, and maintenance.
Open Primary SourceCurtis G. Northcutt, Anish Athalye, Jonas Mueller (NeurIPS Datasets and Benchmarks / JAIR). Demonstrates 3.3% average label corruption across 10 canonical ML benchmarks and formalizes Confident Learning joint noise estimation used in Pillar 02.
Open Primary SourceSamir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Ludwig Schmidt et al. (NeurIPS). Establishes controlled fixed-compute dataset filtering benchmarks proving semantic curation and deduplication dominate raw web crawl scale.
Open Primary SourceMLCommons Data Working Group (Google, Hugging Face, OpenML, ETH Zurich, Harvard). Defines the JSON-LD schema.org/Dataset vocabulary binding physical shard files, record sets, and typed column features across ML loaders.
Open Primary SourceNational Institute of Standards and Technology (U.S. Department of Commerce, NIST AI 100-1). Provides MAP and MEASURE controls governing training data provenance, representativeness, bias auditing, and third-party dataset risk.
Open Primary SourceHugging Face Open-Source Engineering. Canonical technical documentation for zero-copy Arrow memory mapping, Parquet shard streaming, and Hub dataset card metadata schemas.
Open Primary SourceHumanSignal Open-Source Engineering. Reference specification for configuring deterministic multi-rater interfaces across text spans, audio diarization, and frame-indexed video.
Open Primary SourceCleanlab Open-Source Maintainers. Implementation reference for running out-of-fold Confident Learning, near-duplicate detection, and outlier scoring over production datasets.
Open Primary SourceThe Best AI Data Registry restricts primary citations to peer-reviewed conference or journal publications (NeurIPS, ICML, ICLR, JAIR, CACM), recognized international standards consortia (MLCommons, Linux Foundation SPDX, W3C, NIST), and canonical Apache/AGPL open-source documentation with reproducible reference code.
Vendor whitepapers lacking reproducible evaluation harnesses or public mathematical definitions are excluded from normative methodology thresholds.
Academic papers frequently publish theoretical metrics in isolation from storage formats and legal provenance. Our downloadable registry artifacts (/artifacts/modality-qa-thresholds.json and /artifacts/handoff-record.schema.json) unify Gebru et al.'s qualitative Datasheet questions, Northcutt et al.'s quantitative noise ceilings, and MLCommons Croissant structural descriptors into executable JSON Schema validation contracts.