CroissantMiner

A benchmark for extracting Croissant metadata from ML dataset papers, released with the NeurIPS 2026 paper CroissantMiner: Automated Extraction and Validation of Croissant Metadata for ML Datasets (Evaluations and Datasets track).

Code: github.com/berkearda/croissantminer · Leaderboard: berkearda.github.io/croissantminer · Demo: Hugging Face Space · Paper: arXiv link follows · Version reviewed at NeurIPS: croissantminer/croissantminer (tag v1.0)

Contents

Config	Rows	What it is
gold	3,060	Human-validated gold values for 102 datasets x 30 Croissant fields (10 core, 20 Responsible AI). A field the paper does not report is [NULL - not found in paper]; gold_method says how each value was settled.
ratings	10,302	Raw ratings, one row per dataset, field and annotator (pseudonyms A01 to A22); 9,595 after deduplication
calibration	267	Ratings from the calibration phase before the main collection
iaa	30	Per-field agreement: Krippendorff's alpha, Gwet's AC1, raw agreement, bootstrap confidence intervals
annotators	22	Pseudonymized annotator roster: coarse role, ML familiarity, paper experience, hours
silver	500	Paper-level metadata of the 500 silver papers
silver_annotations	15,000	LLM-generated annotations (Claude Sonnet 4.5) for the 500 silver papers x 30 fields; not human-validated
system_outputs	74,400	Outputs of the 24 extraction systems ranked in the paper and of the Claude Sonnet 4.5 reference, on the 14 development and 88 test datasets
judge_verdicts	27,793	GLM-5 judge verdicts on the Responsible AI fields behind the paper's results (1 correct, 2 partially correct, 3 wrong), with the judge's reason
judge_selection	60	Judge selection: human consensus and the scores of the six candidate judges on 30 calibration and 30 validation cells
judge_audit_200	200	Human audit of the judge: ratings by two authors (R1, R2), the third author's rating where they disagreed (R3), the consensus and the judge's score
pilot_gold, pilot_ratings	450, 1,080	Re-annotation pilot with a GPT-5.4 pre-fill, used to check the effect of the seed model; in_paper_analysis marks the 10 papers analysed in the paper

Loading

from datasets import load_dataset

gold = load_dataset("bearda/croissantminer", "gold", split="train")
outputs = load_dataset("bearda/croissantminer", "system_outputs", split="train")

Reproducing the paper

These tables are enough to recompute Table 2 (the composite score of every system), the judge selection table and the judge audit numbers. The code at github.com/berkearda/croissantminer does this (make reproduce, make table2) and explains the scoring rules.

How the gold was made

Claude Sonnet 4.5 pre-filled all 30 fields for each paper. 22 annotators, 8 of them authors of the paper, rated each pre-filled value as correct, partially correct or wrong and corrected it where needed, with at least three ratings per cell. Gold values come from majority vote; cells without a majority were adjudicated, and a later audit corrected 191 cells. Because the pre-fills came from Claude Sonnet 4.5, the paper reports that model as a reference and does not rank it.

Privacy

Annotators appear only as pseudonyms (A01 to A22); names, e-mail addresses and affiliations are not released. The judge-audit raters are three of the authors, labelled R1 to R3. Pilot raters appear only as slots A, B and C per paper.

Versions

• v1.2 (1 October 2026): in gold, the adjudicator ids are the coded labels of v1.0 again (A_senior_01, A_senior_02); all values are unchanged.
• v1.1 (28 September 2026): added silver_annotations, system_outputs, judge_verdicts, judge_selection, judge_audit_200, pilot_gold and pilot_ratings; removed the affiliation column from annotators (it held only a placeholder); gold labels the 14 cells settled by one of the authors adjudicated_berke, the label the code release selects (v1.0 called it adjudicated_senior_02), with identical values; the Croissant file names the authors and corrects a few descriptions.
• v1.0 (7 May 2026): the version reviewed at NeurIPS 2026, also kept unchanged at croissantminer/croissantminer.

The Croissant file (croissant.json) includes the Responsible AI fields required by the NeurIPS 2026 Evaluations and Datasets hosting guidelines.

Limitations

• Gold values record only what the paper states: a field the paper does not report is marked as not found, even when the information exists elsewhere, such as on a dataset card.
• Several Responsible AI fields have few documented cells in the test split, so results for those fields are noisy, and pretraining contamination cannot be ruled out.
• The gold was seeded by Claude Sonnet 4.5, so some anchoring to the pre-fill model cannot be ruled out; the pilot checks this with a GPT-5.4 pre-fill.
• The benchmark covers English-language ML dataset papers and the 30-field Croissant 1.1 schema.

License

The annotations are released under CC BY 4.0. System outputs and judge verdicts quote passages from the benchmark papers, which remain under their authors' licenses, and model outputs are subject to the providers' terms. The code is MIT-licensed on GitHub. Please report problems as GitHub issues.

Citation

@inproceedings{arda2026croissantminer,
  title     = {CroissantMiner: Automated Extraction and Validation of Croissant Metadata for ML Datasets},
  author    = {Arda, Berke and Yavuz, Ahmetcan and Gerry, Paul and Lobentanzer, Sebastian and
               Sarwar, Nobin and Giner-Miguelez, Joan and Chen, Kongtao and Zhang, Luyao and
               Sachan, Mrinmaya and Akhtar, Mubashara},
  booktitle = {Advances in Neural Information Processing Systems (Evaluations and Datasets Track)},
  year      = {2026}
}