Dataset Viewer
Auto-converted to Parquet Duplicate
annotator_id
stringlengths
3
3
n_ratings
int64
12
1.12k
phases_active
stringclasses
9 values
role
stringclasses
6 values
ml_familiarity
stringclasses
3 values
paper_experience
stringclasses
3 values
hours_dedicated
stringclasses
4 values
registered
bool
2 classes
A01
1,123
final_sweep_208,phase2,redistribution,round3
Master's Student
Somewhat familiar - I have used ML datasets in my research
Yes, occasionally
9+ hours
true
A02
65
phase2
Industry / Independent Researcher
Somewhat familiar - I have used ML datasets in my research
Yes, frequently
3-4 hours
true
A03
98
redistribution
Postdoctoral Researcher
Very familiar - I regularly work with ML datasets and their metadata
Yes, frequently
5-8 hours
true
A04
12
phase2
Industry / Independent Researcher
Somewhat familiar - I have used ML datasets in my research
Yes, occasionally
5-8 hours
true
A05
633
final_sweep_28,phase2
PhD Student
Somewhat familiar - I have used ML datasets in my research
Yes, occasionally
3-4 hours
true
A06
68
final_sweep_208,orphan_topup
PhD Student
Very familiar - I regularly work with ML datasets and their metadata
Yes, frequently
10+
false
A07
207
phase2,round3
Master's Student
Somewhat familiar - I have used ML datasets in my research
Yes, occasionally
3-4 hours
true
A08
202
phase2
Industry / Independent Researcher
Basic familiarity - I understand ML concepts but limited hands-on experience with datasets
Rarely
3-4 hours
true
A09
661
phase2,redistribution
Postdoctoral Researcher
Very familiar - I regularly work with ML datasets and their metadata
Yes, frequently
3-4 hours
true
A10
694
phase2,redistribution,round3
Industry / Independent Researcher
Somewhat familiar - I have used ML datasets in my research
Yes, occasionally
5-8 hours
true
A11
306
phase2
Faculty / Professor
Somewhat familiar - I have used ML datasets in my research
Yes, frequently
3-4 hours
true
A12
625
phase2,round3
Industry / Independent Researcher
Very familiar - I regularly work with ML datasets and their metadata
Yes, frequently
3-4 hours
true
A13
612
phase2
Faculty / Professor
Very familiar - I regularly work with ML datasets and their metadata
Yes, frequently
3-4 hours
true
A14
508
phase2,redistribution
Industry / Independent Researcher
Basic familiarity - I understand ML concepts but limited hands-on experience with datasets
Rarely
3-4 hours
true
A15
414
phase2,redistribution
Researcher
Very familiar - I regularly work with ML datasets and their metadata
Yes, frequently
3-4 hours
true
A16
918
phase2,redistribution
PhD Student
Very familiar - I regularly work with ML datasets and their metadata
Yes, frequently
5-8 hours
true
A17
614
dataCollectionType_pass,redistribution,round3
Researcher
null
null
null
false
A18
291
phase2
Faculty / Professor
Very familiar - I regularly work with ML datasets and their metadata
Yes, frequently
9+ hours
true
A19
46
phase2
Master's Student
Very familiar - I regularly work with ML datasets and their metadata
Rarely
3-4 hours
true
A20
711
phase2,round3
Researcher
Somewhat familiar - I have used ML datasets in my research
Yes, frequently
9+ hours
true
A21
631
phase2,round3
Industry / Independent Researcher
Somewhat familiar - I have used ML datasets in my research
Yes, frequently
3-4 hours
true
A22
156
phase2
Faculty / Professor
Somewhat familiar - I have used ML datasets in my research
Yes, occasionally
5-8 hours
true

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:2610.07132 · 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 marked as missing (see Missing values); gold_method says how each value was settled (see Gold labels).
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")

Splits

The 102 gold datasets are split into 14 development and 88 test datasets, and the paper's results use the 88 test datasets. The split is the split column of system_outputs and the file data/agentic/dev_test_split.json in the code repository. Every config here has a single Hugging Face split called train; this is only the default name and does not mean the data is meant for training. Some gold datasets are different Hugging Face datasets described by the same paper; croissant.json lists the source paper of every gold and silver dataset (prov:wasDerivedFrom).

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.

Gold labels

gold_method Cells Meaning
unanimous_3of3 1,623 All three ratings agreed
unanimous_4of4 212 All four ratings agreed
majority_2of3 788 Two of three ratings agreed
majority_3of4 114 Three of four ratings agreed
adjudicated 118 No majority; a senior author settled the cell (adjudicator_id A_senior_01)
audit_corrected_2026-05-04 191 Corrected in the end-to-end author audit; adjudication_notes gives the reason and adjudicator_id the audit round

Another 14 cells, whose last ratings arrived too late for the majority vote, were settled by a senior author (adjudicator_id A_senior_02).

A cell corrected in the audit carries the audit label instead of its earlier label. For this reason the counts above differ from the counts of the annotation process in the paper, where a senior author adjudicated 140 cells.

Missing values

A gold value of [NULL - not found in paper] or [NULL] means the paper does not report the field. A few cells use other forms, such as Not specified, and six cells have an empty value (null); all six were settled by a senior author (A_senior_02). The scoring code treats all of these as not reported (is_null_gold in scripts/figures/build_test88_headline_table.py), which gives 847 of the 3,060 cells (27.7%).

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. Values quote the public source papers, so they contain those papers' author names and a few contact e-mail addresses printed in the papers.

Versions

The main branch is v1.2, the version the paper's results use. Each version is a Git tag and can be loaded with revision="v1.2" (or another tag). Later corrections to the data are released as new tagged versions.

  • Documentation update (5 November 2026): this card and the descriptions in croissant.json (version, maintenance plan, personal information, source papers and provenance) were corrected against the original records; the data files are unchanged and are still v1.2.
  • 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 gives the 14 cells settled by one of the authors 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}
}
Downloads last month
581

Space using bearda/croissantminer 1

Paper for bearda/croissantminer