Research map / Large language models / AI and society

Croissant Baker: Metadata Generation for Discoverable, Governable, and Reusable ML Datasets

NeurIPS 2026 · Large language models › AI and society · approach: Benchmarking and reproducibility practice · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Releases Croissant Baker, a local-first open-source CLI that deterministically generates validated Croissant JSON-LD metadata from dataset directories via a modular format-handler registry, evaluated on 140+ datasets and several held-out splits.

Problem

Croissant metadata generation on existing platforms requires uploading data, which is infeasible for governed or large local repositories, and generic file walks and platform generators misrecover structure in scientific and domain-specific formats (WFDB, FHIR, DICOM, NIfTI, partitioned Parquet, mul

Approach

A local-first, deterministic structural recovery layer with a typed handler registry for heterogeneous scientific formats, separated from optional semantic enrichment supplied by CLI flags or agents; outputs are traceable to file bytes or explicit inputs.

Key results

• Semantic type agreement of 97.9% on 235 fields across a seeded 25-dataset NeurIPS 2025 D&B cross-domain sample (24/25 produce valid Croissant).
• 97.4% semantic type agreement (798/819 fields) on 55 Open Targets producer-authored Parquet Croissant datasets, with 55/55 RecordSet names and 819/819 field names recovered.
• Scales to MIMIC-IV full (9.92 GB) in 13.3 s and MIMIC-IV MEDS full (374 Parquet files, 3.67 GB) in 32.2 s, with all nine local outputs passing mlcroissant validation.

Surprising findings

• Most Open Targets and FHIR type disagreements stem from producer Croissants using sc:Float or sc:Text where Baker infers more precise integer or URL types, so the reference itself is sometimes less precise.
• Platform auto-generators (HF, Kaggle) misclassify integer IDs as text and datetimes as dates, and produce empty schemas for WFDB and multi-band TIFF.

Limitations

• Type inference uses heuristics on sampled rows and may misclassify ambiguous fields.
• Unsupported formats (e.g., Apache Arrow in UniHG) are skipped, and inter-table relationships such as foreign keys are not inferred.
• Semantic fields (description, citation, license, creator) must be supplied by humans or agents; evaluation is limited to a few format families and most held-out ground truths are producer-authored or standards-derived.

Experimental setup

Datasets: MIMIC-IV demo MIMIC-IV full (CSV) MIMIC-IV MEDS eICU-CRD demo MIT-BIH Arrhythmia MIMIC-IV OMOP Glaucoma Fundus (HYGD) Satellite Public Health NeurIPS 2025 D&B datasets (25 sampled) Open Targets (55 Parquet datasets) SMART Health IT FHIR dcm_validate corpus

Benchmarks: MIMIC-IV full (CSV) MIMIC-IV MEDS full NeurIPS 2025 D&B cross-domain sample Open Targets SMART Health IT FHIR dcm_validate DICOM PS3.6 OpenNeuro BIDS

Metrics: semantic type agreement strict type agreement field-name recovery RecordSet-name recovery mlcroissant validity execution runtime

Baselines: Hugging Face auto-generated Croissant Kaggle auto-generated Croissant Croissant Editor manual authoring

Builds on: Croissant mlcroissant Schema.org Model Context Protocol

Closest papers by meaning

• Croissant: A Metadata Format for ML-Ready Datasets (NeurIPS 2024 · similarity 0.96)
• CroissantMiner: Automated Extraction and Validation of Croissant Metadata for ML Datasets (NeurIPS 2026 · similarity 0.94)
• The State of Data Curation at NeurIPS: An Assessment of Dataset Development Practices in the Datasets and Benchmarks Track (NeurIPS 2024 · similarity 0.84)
• Bee: A High-Quality Corpus and Full-Stack Suite to Unlock Advanced Fully Open MLLMs (ICLR 2026 · similarity 0.82)
• STAMP Your Content: Proving Dataset Membership via Watermarked Rephrasings (ICML 2025 · similarity 0.81)
• Navigating Dataset Documentations in AI: A Large-Scale Analysis of Dataset Cards on HuggingFace (ICLR 2024 · similarity 0.81)