CroissantMiner v1.0: the version reviewed for NeurIPS 2026 (unchanged copy of croissantminer/croissantminer)

README.md

@@ -0,0 +1,136 @@
+---
+license: cc-by-4.0
+language:
+- en
+tags:
+- metadata
+- responsible-ai
+- benchmark
+- croissant
+- ml-datasets
+size_categories:
+- n<1K
+configs:
+- config_name: ratings
+  data_files:
+  - split: train
+    path: ratings.parquet
+- config_name: gold
+  data_files:
+  - split: train
+    path: gold.parquet
+- config_name: silver
+  data_files:
+  - split: train
+    path: silver.parquet
+- config_name: calibration
+  data_files:
+  - split: train
+    path: calibration.parquet
+- config_name: annotators
+  data_files:
+  - split: train
+    path: annotators.parquet
+- config_name: iaa
+  data_files:
+  - split: train
+    path: iaa.parquet
+task_categories:
+- text-classification
+- text-generation
+pretty_name: CroissantMiner
+---
+
+# CroissantMiner: A Benchmark for Automated Croissant Metadata Extraction
+
+CroissantMiner is a benchmark and evaluation suite for automated extraction of MLCommons Croissant 1.1 metadata from machine-learning dataset papers. The headline gold split pairs **102 papers** with the full **30-field Croissant 1.1 schema** (3,060 human-validated cells); a 500-paper silver split adds paper-level metadata for scale studies. The release supports evaluation of single-pass, fine-tuned, and agentic extraction systems.
+
+## Splits
+
+| Split | Papers | Coverage | Annotation |
+|---|---:|---|---|
+| Gold | 102 | full 30-field schema (3,060 cells) | 22-annotator human-validated, majority vote with senior adjudication, end-to-end author audit |
+| Silver | 500 | paper-level metadata only — 9 fields (downloads, domain, license, tasks, language, datePublished, dataCollectionType, arxiv_id, paper_id) | LLM-derived from arXiv metadata + Claude Sonnet 4.5 pre-fill |
+| **Total** | **602** | | |
+
+## What's in the Croissant metadata
+
+Each gold cell carries:
+- **Field schema:** 10 General fields (`sc:name`, `sc:license`, `sc:citeAs`, …) + 20 Responsible-AI fields (`rai:dataCollection`, `rai:dataBiases`, `rai:personalSensitiveInformation`, …).
+- **Gold value:** the human-validated string or `[NULL ...]` if the paper does not document the field.
+- **Method:** how the gold value was derived (`unanimous_3of3`, `majority_2of3`, `adjudicated`, `audit_corrected_2026-05-04`).
+- **Provenance:** annotator IDs, rating counts, IAA per field.
+
+## Mandatory Croissant RAI fields (per NeurIPS 2026 E&D)
+
+The Croissant JSON-LD file populates all 8 RAI fields required by the [NeurIPS 2026 Evaluations & Datasets hosting guidelines](https://neurips.cc/Conferences/2026/EvaluationsDatasetsHosting):
+
+- `rai:dataLimitations` — known limitations of the annotation rubric and source-paper sample
+- `rai:dataBiases` — annotator-pool, source-paper-domain, and field-difficulty biases
+- `rai:personalSensitiveInformation` — none; ratings are aggregate, no per-annotator demographic data
+- `rai:dataUseCases` — Training / Testing / Validation / Benchmarking of LLM-based metadata extraction
+- `rai:dataSocialImpact` — opens RAI-field difficulty research; reduces manual metadata burden
+- `rai:isSyntheticData` — `False`; ratings are produced on real LLM-generated extractions of real published papers
+- `rai:sourceDatasets` — public Hugging Face datasets ranked by download count, paper PDFs from arXiv
+- `rai:provenanceActivities` — full pipeline: Sonnet 4.5 pre-fill → 22-annotator review → majority vote → senior adjudication → 2026-05-04 author audit
+
+## Loading
+
+```python
+from datasets import load_dataset
+
+# Gold values: human-validated (paper, field) cells
+gold = load_dataset("croissantminer/croissantminer", "gold")
+
+# Silver paper-level metadata: 500 papers, 9 fields
+silver = load_dataset("croissantminer/croissantminer", "silver")
+
+# Raw long-format ratings: one row per (paper, field, annotator)
+ratings = load_dataset("croissantminer/croissantminer", "ratings")
+
+# Per-field inter-annotator agreement (alpha + AC1 + bootstrap CIs)
+iaa = load_dataset("croissantminer/croissantminer", "iaa")
+
+# Pseudonymized annotator roster (A01..A22) + aggregate demographics
+annotators = load_dataset("croissantminer/croissantminer", "annotators")
+```
+
+## License
+
+- **Benchmark data and metadata**: CC-BY-4.0
+- **Tool / extraction code**: Apache-2.0 (released separately at the GitHub repository)
+
+## Citation
+
+If you use CroissantMiner, please cite the paper:
+
+```bibtex
+@inproceedings{croissantminer2026,
+  title={CroissantMiner: Automated Extraction and Validation of Croissant Metadata for ML Datasets},
+  author={[Anonymized for review]},
+  booktitle={NeurIPS 2026 Evaluations and Datasets Track},
+  year={2026}
+}
+```
+
+## Limitations
+
+See the paper §7 and Appendix~I for the full discussion. Headline limitations:
+
+- **PDF parser noise:** PyPDF2 misses figure/table content where some RAI fields' evidence lives.
+- **English-only gold:** all 102 gold papers are in English.
+- **Benchmark scale:** the 102-paper gold split is too small for tight per-domain stratification.
+- **Gold-seed dependency:** gold values were derived from human verification of Claude Sonnet 4.5 pre-fills; structural choices about what counts as "filled" are inherited from the seed model.
+
+## Reproducibility
+
+Reproduction instructions for Table 2 are in the anonymized code repository's README under `Reproduction (Table 2)`: install with `pip install -e ".[hf]"`, pull `gold.parquet` from this HuggingFace dataset, run an extraction (`scripts/experiments/model_comparison/extract_all_models_batch.py`, `scripts/agentic_v2.py`, or `scripts/agentic_lev.py`), then score with `evaluation/score_against_gold.py`. Reproducing only the leader row is under \$5; the full lineup is ~\$80 in commercial-API tokens.
+
+## Hosting
+
+- **Long-term hosting:** This dataset commits to long-term availability on Hugging Face. If access changes are needed (e.g., for credentialised access), the maintainers will update the dataset card and notify NeurIPS.
+- **DOI:** TBD (will be assigned at camera-ready).
+
+## Maintenance
+
+The dataset is versioned by date (current: `v1.0.0`, 2026-05-07). Updates with re-runs against newer model releases will be tagged as `v1.x`; structural schema changes (e.g., Croissant 2.x migration) will be tagged as `v2.0`.

annotators.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f9c107a576b569a00bb92a2579ada8b53abd1834e3d25dd409f668a0c98287f2
+size 6868

calibration.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:669e670951f8514d23603caccfd6f230bce2919bf20edc58425fbdf4a1ba8b3e
+size 42548

croissant.json

@@ -0,0 +1,1178 @@
+{
+  "@context": {
+    "@language": "en",
+    "@vocab": "https://schema.org/",
+    "citeAs": "cr:citeAs",
+    "column": "cr:column",
+    "conformsTo": "dct:conformsTo",
+    "cr": "http://mlcommons.org/croissant/",
+    "rai": "http://mlcommons.org/croissant/RAI/",
+    "prov": "http://www.w3.org/ns/prov#",
+    "data": {
+      "@id": "cr:data",
+      "@type": "@json"
+    },
+    "dataType": {
+      "@id": "cr:dataType",
+      "@type": "@vocab"
+    },
+    "dct": "http://purl.org/dc/terms/",
+    "examples": {
+      "@id": "cr:examples",
+      "@type": "@json"
+    },
+    "extract": "cr:extract",
+    "field": "cr:field",
+    "fileObject": "cr:fileObject",
+    "fileSet": "cr:fileSet",
+    "format": "cr:format",
+    "includes": "cr:includes",
+    "isLiveDataset": "cr:isLiveDataset",
+    "jsonPath": "cr:jsonPath",
+    "key": "cr:key",
+    "md5": "cr:md5",
+    "parentField": "cr:parentField",
+    "path": "cr:path",
+    "recordSet": "cr:recordSet",
+    "references": "cr:references",
+    "regex": "cr:regex",
+    "repeated": "cr:repeated",
+    "replace": "cr:replace",
+    "sc": "https://schema.org/",
+    "separator": "cr:separator",
+    "source": "cr:source",
+    "subField": "cr:subField",
+    "transform": "cr:transform"
+  },
+  "@type": "sc:Dataset",
+  "@id": "https://anonymous.4open.science/r/croissantminer-F87C/tree/main/data/annotations",
+  "conformsTo": "http://mlcommons.org/croissant/1.1",
+  "name": "CroissantMiner Annotations",
+  "description": "Human-validated annotation dataset accompanying the CroissantMiner benchmark for automated extraction of MLCommons Croissant Responsible AI metadata from ML dataset papers. Contains 9595 ratings and 267 calibration ratings from 22 annotators across 102 papers and 30 Croissant Core + RAI fields, with derived gold values, per-field inter-annotator agreement metrics, and pseudonymized annotator demographics. Released under CC-BY-4.0; LLM-generated extractions referenced by the ratings are subject to their respective providers' terms.",
+  "url": "https://anonymous.4open.science/r/croissantminer-F87C",
+  "version": "1.0.0",
+  "datePublished": "2026-05-07",
+  "license": "https://creativecommons.org/licenses/by/4.0/",
+  "inLanguage": "en",
+  "isLiveDataset": false,
+  "keywords": [
+    "Responsible AI",
+    "metadata extraction",
+    "Croissant",
+    "LLM evaluation",
+    "human annotation",
+    "inter-annotator agreement",
+    "benchmark"
+  ],
+  "creator": [
+    {
+      "@type": "sc:Person",
+      "name": "Anonymous Author(s)",
+      "affiliation": "Anonymous"
+    }
+  ],
+  "publisher": {
+    "@type": "sc:Organization",
+    "name": "Anonymous"
+  },
+  "citeAs": "Anonymous Authors. CroissantMiner: Automated Extraction and Validation of Croissant Metadata for ML Datasets. NeurIPS 2026 Evaluations and Datasets Track (under review).",
+  "distribution": [
+    {
+      "@type": "cr:FileObject",
+      "@id": "ratings-parquet",
+      "name": "ratings.parquet",
+      "description": "Long-format raw human ratings, one row per (paper, field, annotator). Annotators were instructed to rate each LLM-extracted Croissant metadata value as Correct (1), Partially Correct (2), or Not Correct (3); for ratings of 2 or 3, an optional corrected_value and free-text notes were collected, plus a confidence label (High / Medium / Low) and a failure-mode category.",
+      "contentUrl": "https://huggingface.co/datasets/croissantminer/croissantminer/resolve/main/ratings.parquet",
+      "encodingFormat": "application/x-parquet",
+      "sha256": "0eec928fba8da7df52ab87da403f2448189106252c35222da05d2553c6ed5f1f"
+    },
+    {
+      "@type": "cr:FileObject",
+      "@id": "calibration-parquet",
+      "name": "calibration.parquet",
+      "description": "Pilot ratings collected before the main collection began. Used a binary TRUE/FALSE rubric (mapped to rating 1 or 3 in this file) on a 15-cell sample. Released for transparency; downstream evaluation should use ratings.parquet.",
+      "contentUrl": "https://huggingface.co/datasets/croissantminer/croissantminer/resolve/main/calibration.parquet",
+      "encodingFormat": "application/x-parquet",
+      "sha256": "669e670951f8514d23603caccfd6f230bce2919bf20edc58425fbdf4a1ba8b3e"
+    },
+    {
+      "@type": "cr:FileObject",
+      "@id": "gold-parquet",
+      "name": "gold.parquet",
+      "description": "Derived gold values, one row per (paper, field). Computed by majority vote across raters (3 to 4 raters per cell, see n_raters column); cells where no rating reached majority (gold_method = 'tie_split') were sent to senior adjudication and the gold_value/gold_rating were filled by the adjudicator.",
+      "contentUrl": "https://huggingface.co/datasets/croissantminer/croissantminer/resolve/main/gold.parquet",
+      "encodingFormat": "application/x-parquet",
+      "sha256": "b510bf7a81a163a434ac5ec55af0c136c5e97da6c2e0e3807e7b0270bdde8fcd"
+    },
+    {
+      "@type": "cr:FileObject",
+      "@id": "iaa-parquet",
+      "name": "iaa.parquet",
+      "description": "Per-field inter-annotator agreement metrics with bootstrap 95 percent confidence intervals. Reports Krippendorff's alpha (ordinal), Gwet's AC1 (chance-corrected, prevalence-robust), raw pairwise agreement, and percent unanimous cells. The AC1 + pairwise pair triangulates against the prevalence paradox in alpha when ratings cluster on one category.",
+      "contentUrl": "https://huggingface.co/datasets/croissantminer/croissantminer/resolve/main/iaa.parquet",
+      "encodingFormat": "application/x-parquet",
+      "sha256": "affcab7689b48213dce2de0a2a2ad686805d2b0f04a51f23f8079afdeea43520"
+    },
+    {
+      "@type": "cr:FileObject",
+      "@id": "annotators-parquet",
+      "name": "annotators.parquet",
+      "description": "Pseudonymized roster of the 22 annotators who contributed to the main set. Demographics aggregated from the recruitment registration form: role, affiliation, ML familiarity, paper-experience frequency, hours dedicated. No personally identifying information beyond what annotators publicly associate with their professional profiles.",
+      "contentUrl": "https://huggingface.co/datasets/croissantminer/croissantminer/resolve/main/annotators.parquet",
+      "encodingFormat": "application/x-parquet",
+      "sha256": "f9c107a576b569a00bb92a2579ada8b53abd1834e3d25dd409f668a0c98287f2"
+    },
+    {
+      "@type": "cr:FileObject",
+      "@id": "silver-parquet",
+      "name": "silver.parquet",
+      "description": "Silver split: 500 ML dataset papers with paper-level metadata (arXiv ID, downloads, domain, license, tasks, language, datePublished, dataCollectionType). LLM-derived from arXiv metadata + Claude Sonnet 4.5 pre-fill at temperature 0; not human-validated. Used for scale studies and silver-vs-gold validation in the paper.",
+      "contentUrl": "https://huggingface.co/datasets/croissantminer/croissantminer/resolve/main/silver.parquet",
+      "encodingFormat": "application/x-parquet",
+      "sha256": "5b77b10bf683e2c05b1a03887f0ab92adebf595df7d78d28dacf8cea30e2f581"
+    }
+  ],
+  "recordSet": [
+    {
+      "@type": "cr:RecordSet",
+      "@id": "ratings",
+      "name": "ratings",
+      "description": "Long-format raw human ratings, one row per (paper, field, annotator). Annotators were instructed to rate each LLM-extracted Croissant metadata value as Correct (1), Partially Correct (2), or Not Correct (3); for ratings of 2 or 3, an optional corrected_value and free-text notes were collected, plus a confidence label (High / Medium / Low) and a failure-mode category.",
+      "field": [
+        {
+          "@type": "cr:Field",
+          "@id": "ratings/paper_id",
+          "name": "paper_id",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {
+              "@id": "ratings-parquet"
+            },
+            "extract": {
+              "column": "paper_id"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "ratings/field_id",
+          "name": "field_id",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {
+              "@id": "ratings-parquet"
+            },
+            "extract": {
+              "column": "field_id"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "ratings/annotator",
+          "name": "annotator",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {
+              "@id": "ratings-parquet"
+            },
+            "extract": {
+              "column": "annotator"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "ratings/phase",
+          "name": "phase",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {
+              "@id": "ratings-parquet"
+            },
+            "extract": {
+              "column": "phase"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "ratings/rating",
+          "name": "rating",
+          "dataType": "sc:Integer",
+          "source": {
+            "fileObject": {
+              "@id": "ratings-parquet"
+            },
+            "extract": {
+              "column": "rating"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "ratings/extracted_value",
+          "name": "extracted_value",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {
+              "@id": "ratings-parquet"
+            },
+            "extract": {
+              "column": "extracted_value"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "ratings/corrected_value",
+          "name": "corrected_value",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {
+              "@id": "ratings-parquet"
+            },
+            "extract": {
+              "column": "corrected_value"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "ratings/failure_mode",
+          "name": "failure_mode",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {
+              "@id": "ratings-parquet"
+            },
+            "extract": {
+              "column": "failure_mode"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "ratings/confidence",
+          "name": "confidence",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {
+              "@id": "ratings-parquet"
+            },
+            "extract": {
+              "column": "confidence"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "ratings/notes",
+          "name": "notes",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {
+              "@id": "ratings-parquet"
+            },
+            "extract": {
+              "column": "notes"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "ratings/source_file",
+          "name": "source_file",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {
+              "@id": "ratings-parquet"
+            },
+            "extract": {
+              "column": "source_file"
+            }
+          }
+        }
+      ]
+    },
+    {
+      "@type": "cr:RecordSet",
+      "@id": "calibration",
+      "name": "calibration",
+      "description": "Pilot ratings collected before the main collection began. Used a binary TRUE/FALSE rubric (mapped to rating 1 or 3 in this file) on a 15-cell sample. Released for transparency; downstream evaluation should use ratings.parquet.",
+      "field": [
+        {
+          "@type": "cr:Field",
+          "@id": "calibration/paper_id",
+          "name": "paper_id",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {
+              "@id": "calibration-parquet"
+            },
+            "extract": {
+              "column": "paper_id"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "calibration/field_id",
+          "name": "field_id",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {
+              "@id": "calibration-parquet"
+            },
+            "extract": {
+              "column": "field_id"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "calibration/annotator",
+          "name": "annotator",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {
+              "@id": "calibration-parquet"
+            },
+            "extract": {
+              "column": "annotator"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "calibration/phase",
+          "name": "phase",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {
+              "@id": "calibration-parquet"
+            },
+            "extract": {
+              "column": "phase"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "calibration/rating",
+          "name": "rating",
+          "dataType": "sc:Integer",
+          "source": {
+            "fileObject": {
+              "@id": "calibration-parquet"
+            },
+            "extract": {
+              "column": "rating"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "calibration/extracted_value",
+          "name": "extracted_value",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {
+              "@id": "calibration-parquet"
+            },
+            "extract": {
+              "column": "extracted_value"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "calibration/corrected_value",
+          "name": "corrected_value",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {
+              "@id": "calibration-parquet"
+            },
+            "extract": {
+              "column": "corrected_value"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "calibration/failure_mode",
+          "name": "failure_mode",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {
+              "@id": "calibration-parquet"
+            },
+            "extract": {
+              "column": "failure_mode"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "calibration/confidence",
+          "name": "confidence",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {
+              "@id": "calibration-parquet"
+            },
+            "extract": {
+              "column": "confidence"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "calibration/notes",
+          "name": "notes",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {
+              "@id": "calibration-parquet"
+            },
+            "extract": {
+              "column": "notes"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "calibration/source_file",
+          "name": "source_file",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {
+              "@id": "calibration-parquet"
+            },
+            "extract": {
+              "column": "source_file"
+            }
+          }
+        }
+      ]
+    },
+    {
+      "@type": "cr:RecordSet",
+      "@id": "gold",
+      "name": "gold",
+      "description": "Derived gold values, one row per (paper, field). Computed by majority vote across raters (3 to 4 raters per cell, see n_raters column); cells where no rating reached majority (gold_method = 'tie_split') were sent to senior adjudication and the gold_value/gold_rating were filled by the adjudicator.",
+      "field": [
+        {
+          "@type": "cr:Field",
+          "@id": "gold/paper_id",
+          "name": "paper_id",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {
+              "@id": "gold-parquet"
+            },
+            "extract": {
+              "column": "paper_id"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "gold/field_id",
+          "name": "field_id",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {
+              "@id": "gold-parquet"
+            },
+            "extract": {
+              "column": "field_id"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "gold/gold_value",
+          "name": "gold_value",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {
+              "@id": "gold-parquet"
+            },
+            "extract": {
+              "column": "gold_value"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "gold/gold_rating",
+          "name": "gold_rating",
+          "dataType": "sc:Float",
+          "source": {
+            "fileObject": {
+              "@id": "gold-parquet"
+            },
+            "extract": {
+              "column": "gold_rating"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "gold/gold_method",
+          "name": "gold_method",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {
+              "@id": "gold-parquet"
+            },
+            "extract": {
+              "column": "gold_method"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "gold/n_raters",
+          "name": "n_raters",
+          "dataType": "sc:Integer",
+          "source": {
+            "fileObject": {
+              "@id": "gold-parquet"
+            },
+            "extract": {
+              "column": "n_raters"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "gold/n_agree",
+          "name": "n_agree",
+          "dataType": "sc:Integer",
+          "source": {
+            "fileObject": {
+              "@id": "gold-parquet"
+            },
+            "extract": {
+              "column": "n_agree"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "gold/consensus_pct",
+          "name": "consensus_pct",
+          "dataType": "sc:Float",
+          "source": {
+            "fileObject": {
+              "@id": "gold-parquet"
+            },
+            "extract": {
+              "column": "consensus_pct"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "gold/adjudicator_id",
+          "name": "adjudicator_id",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {
+              "@id": "gold-parquet"
+            },
+            "extract": {
+              "column": "adjudicator_id"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "gold/adjudication_notes",
+          "name": "adjudication_notes",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {
+              "@id": "gold-parquet"
+            },
+            "extract": {
+              "column": "adjudication_notes"
+            }
+          }
+        }
+      ]
+    },
+    {
+      "@type": "cr:RecordSet",
+      "@id": "iaa",
+      "name": "iaa",
+      "description": "Per-field inter-annotator agreement metrics with bootstrap 95 percent confidence intervals. Reports Krippendorff's alpha (ordinal), Gwet's AC1 (chance-corrected, prevalence-robust), raw pairwise agreement, and percent unanimous cells. The AC1 + pairwise pair triangulates against the prevalence paradox in alpha when ratings cluster on one category.",
+      "field": [
+        {
+          "@type": "cr:Field",
+          "@id": "iaa/field_id",
+          "name": "field_id",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {
+              "@id": "iaa-parquet"
+            },
+            "extract": {
+              "column": "field_id"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "iaa/level",
+          "name": "level",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {
+              "@id": "iaa-parquet"
+            },
+            "extract": {
+              "column": "level"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "iaa/alpha",
+          "name": "alpha",
+          "dataType": "sc:Float",
+          "source": {
+            "fileObject": {
+              "@id": "iaa-parquet"
+            },
+            "extract": {
+              "column": "alpha"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "iaa/alpha_ci_low",
+          "name": "alpha_ci_low",
+          "dataType": "sc:Float",
+          "source": {
+            "fileObject": {
+              "@id": "iaa-parquet"
+            },
+            "extract": {
+              "column": "alpha_ci_low"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "iaa/alpha_ci_high",
+          "name": "alpha_ci_high",
+          "dataType": "sc:Float",
+          "source": {
+            "fileObject": {
+              "@id": "iaa-parquet"
+            },
+            "extract": {
+              "column": "alpha_ci_high"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "iaa/gwet_ac1",
+          "name": "gwet_ac1",
+          "dataType": "sc:Float",
+          "source": {
+            "fileObject": {
+              "@id": "iaa-parquet"
+            },
+            "extract": {
+              "column": "gwet_ac1"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "iaa/gwet_ac1_ci_low",
+          "name": "gwet_ac1_ci_low",
+          "dataType": "sc:Float",
+          "source": {
+            "fileObject": {
+              "@id": "iaa-parquet"
+            },
+            "extract": {
+              "column": "gwet_ac1_ci_low"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "iaa/gwet_ac1_ci_high",
+          "name": "gwet_ac1_ci_high",
+          "dataType": "sc:Float",
+          "source": {
+            "fileObject": {
+              "@id": "iaa-parquet"
+            },
+            "extract": {
+              "column": "gwet_ac1_ci_high"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "iaa/pct_pairwise_agreement",
+          "name": "pct_pairwise_agreement",
+          "dataType": "sc:Float",
+          "source": {
+            "fileObject": {
+              "@id": "iaa-parquet"
+            },
+            "extract": {
+              "column": "pct_pairwise_agreement"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "iaa/pct_unanimous",
+          "name": "pct_unanimous",
+          "dataType": "sc:Float",
+          "source": {
+            "fileObject": {
+              "@id": "iaa-parquet"
+            },
+            "extract": {
+              "column": "pct_unanimous"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "iaa/n_unanimous_cells",
+          "name": "n_unanimous_cells",
+          "dataType": "sc:Integer",
+          "source": {
+            "fileObject": {
+              "@id": "iaa-parquet"
+            },
+            "extract": {
+              "column": "n_unanimous_cells"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "iaa/n_cells_with_2plus_raters",
+          "name": "n_cells_with_2plus_raters",
+          "dataType": "sc:Integer",
+          "source": {
+            "fileObject": {
+              "@id": "iaa-parquet"
+            },
+            "extract": {
+              "column": "n_cells_with_2plus_raters"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "iaa/n_papers",
+          "name": "n_papers",
+          "dataType": "sc:Integer",
+          "source": {
+            "fileObject": {
+              "@id": "iaa-parquet"
+            },
+            "extract": {
+              "column": "n_papers"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "iaa/n_active_annotators",
+          "name": "n_active_annotators",
+          "dataType": "sc:Integer",
+          "source": {
+            "fileObject": {
+              "@id": "iaa-parquet"
+            },
+            "extract": {
+              "column": "n_active_annotators"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "iaa/n_total_ratings",
+          "name": "n_total_ratings",
+          "dataType": "sc:Integer",
+          "source": {
+            "fileObject": {
+              "@id": "iaa-parquet"
+            },
+            "extract": {
+              "column": "n_total_ratings"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "iaa/note",
+          "name": "note",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {
+              "@id": "iaa-parquet"
+            },
+            "extract": {
+              "column": "note"
+            }
+          }
+        }
+      ]
+    },
+    {
+      "@type": "cr:RecordSet",
+      "@id": "annotators",
+      "name": "annotators",
+      "description": "Pseudonymized roster of the 22 annotators who contributed to the main set. Demographics aggregated from the recruitment registration form: role, affiliation, ML familiarity, paper-experience frequency, hours dedicated. No personally identifying information beyond what annotators publicly associate with their professional profiles.",
+      "field": [
+        {
+          "@type": "cr:Field",
+          "@id": "annotators/annotator_id",
+          "name": "annotator_id",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {
+              "@id": "annotators-parquet"
+            },
+            "extract": {
+              "column": "annotator_id"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "annotators/n_ratings",
+          "name": "n_ratings",
+          "dataType": "sc:Integer",
+          "source": {
+            "fileObject": {
+              "@id": "annotators-parquet"
+            },
+            "extract": {
+              "column": "n_ratings"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "annotators/phases_active",
+          "name": "phases_active",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {
+              "@id": "annotators-parquet"
+            },
+            "extract": {
+              "column": "phases_active"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "annotators/role",
+          "name": "role",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {
+              "@id": "annotators-parquet"
+            },
+            "extract": {
+              "column": "role"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "annotators/affiliation",
+          "name": "affiliation",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {
+              "@id": "annotators-parquet"
+            },
+            "extract": {
+              "column": "affiliation"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "annotators/ml_familiarity",
+          "name": "ml_familiarity",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {
+              "@id": "annotators-parquet"
+            },
+            "extract": {
+              "column": "ml_familiarity"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "annotators/paper_experience",
+          "name": "paper_experience",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {
+              "@id": "annotators-parquet"
+            },
+            "extract": {
+              "column": "paper_experience"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "annotators/hours_dedicated",
+          "name": "hours_dedicated",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {
+              "@id": "annotators-parquet"
+            },
+            "extract": {
+              "column": "hours_dedicated"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "annotators/registered",
+          "name": "registered",
+          "dataType": "sc:Boolean",
+          "source": {
+            "fileObject": {
+              "@id": "annotators-parquet"
+            },
+            "extract": {
+              "column": "registered"
+            }
+          }
+        }
+      ]
+    },
+    {
+      "@type": "cr:RecordSet",
+      "@id": "silver",
+      "name": "silver",
+      "description": "Silver split: 500 ML dataset papers with paper-level metadata (arXiv ID, downloads, domain, license, tasks, language, datePublished, dataCollectionType). LLM-derived from arXiv metadata + Claude Sonnet 4.5 pre-fill at temperature 0; not human-validated. Used for scale studies and silver-vs-gold validation in the paper.",
+      "field": [
+        {
+          "@type": "cr:Field",
+          "@id": "silver/paper_id",
+          "name": "paper_id",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {
+              "@id": "silver-parquet"
+            },
+            "extract": {
+              "column": "paper_id"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "silver/arxiv_id",
+          "name": "arxiv_id",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {
+              "@id": "silver-parquet"
+            },
+            "extract": {
+              "column": "arxiv_id"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "silver/downloads",
+          "name": "downloads",
+          "dataType": "sc:Integer",
+          "source": {
+            "fileObject": {
+              "@id": "silver-parquet"
+            },
+            "extract": {
+              "column": "downloads"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "silver/domain",
+          "name": "domain",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {
+              "@id": "silver-parquet"
+            },
+            "extract": {
+              "column": "domain"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "silver/license",
+          "name": "license",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {
+              "@id": "silver-parquet"
+            },
+            "extract": {
+              "column": "license"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "silver/tasks",
+          "name": "tasks",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {
+              "@id": "silver-parquet"
+            },
+            "extract": {
+              "column": "tasks"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "silver/inLanguage",
+          "name": "inLanguage",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {
+              "@id": "silver-parquet"
+            },
+            "extract": {
+              "column": "inLanguage"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "silver/datePublished",
+          "name": "datePublished",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {
+              "@id": "silver-parquet"
+            },
+            "extract": {
+              "column": "datePublished"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "silver/dataCollectionType",
+          "name": "dataCollectionType",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {
+              "@id": "silver-parquet"
+            },
+            "extract": {
+              "column": "dataCollectionType"
+            }
+          }
+        }
+      ]
+    }
+  ],
+  "rai:dataCollection": "Human ratings of LLM-extracted Croissant metadata across 102 ML dataset papers and 30 RAI fields. Collected over seven phases between March and April 2026: pilot calibration, Phase 2 main collection, mid-project redistribution to balance coverage, a round-3 top-up, two final sweeps for cells under the three-rater target, and an orphan top-up to bring every cell to at least three distinct annotators. Senior adjudication resolved cells where no rating reached majority.",
+  "rai:dataCollectionType": "Manual Human Curator. Annotators were instructed not to use AI tools while rating.",
+  "rai:dataCollectionTimeframe": "2026-03-10 to 2026-04-25",
+  "rai:dataCollectionRawData": "AI-extracted Croissant metadata values produced by Claude Sonnet 4.5 on 102 arXiv-published ML dataset papers, plus the papers themselves (read by annotators when validating each value).",
+  "rai:dataAnnotationProtocol": "Each (paper, field) cell received three independent ratings on a three-level rubric (Correct / Partially Correct / Not Correct) with six failure-mode categories (Incomplete, Hallucination, Wrong Section, Granularity Mismatch, Format Error, Other) and an optional corrected value plus free-text notes. Confidence (High / Medium / Low) was recorded per rating.",
+  "rai:dataAnnotationPlatform": "Excel files distributed via shared cloud storage; one workbook per annotator with separate sheets for Instructions, Field Definitions, Calibration Examples, and Annotations.",
+  "rai:dataAnnotationAnalysis": "Krippendorff's alpha (ordinal, primary), Gwet's AC1 (chance-corrected, prevalence-robust), raw pairwise agreement, and percent unanimous cells, all computed per field with 1000-iteration bootstrap 95 percent confidence intervals. Triangulation across alpha and AC1 is necessary because alpha collapses under skewed marginals (Feinstein-Cicchetti paradox).",
+  "rai:annotationsPerItem": "3",
+  "rai:annotatorDemographics": "22 distinct annotators after deduplicating two who registered twice and merging name variants. Roles spanned PhD students, postdoctoral researchers, faculty, and industry professionals across academic and external institutions. Aggregate demographics in annotators.parquet (role, affiliation, ML familiarity tier, paper-experience frequency, hours dedicated). No personally identifying demographic data (age, gender, ethnicity, geography) was collected.",
+  "rai:machineAnnotationTools": "None. Annotators were instructed not to use AI tools while rating. The pre-filled metadata values rated by annotators were generated by Claude Sonnet 4.5 (not part of the rating step itself).",
+  "rai:dataReleaseMaintenancePlan": "Released alongside the CroissantMiner NeurIPS 2026 E&D paper. Maintained at https://anonymous.4open.science/r/croissantminer-F87C. Versioning via Git tags. Bug-fix updates expected through 2027; substantive schema changes deprecated rather than silently overwritten.",
+  "rai:personalSensitiveInformation": "None collected. Annotator names appear in the local raw ratings.parquet for build provenance and are replaced with pseudonymous integer IDs (A01..A22) before public release; the raw-name version is gitignored.",
+  "rai:dataSocialImpact": "Enables study of LLM extraction quality on Responsible AI metadata. Permits downstream research into RAI-field difficulty, extraction-distillation, and disagreement-aware fine-tuning. Limitation: source papers are arXiv-published English-language ML dataset papers, so coverage is biased toward the English-speaking ML research community.",
+  "rai:dataBiases": "Annotator pool is biased toward ML/AI researchers (recruited via an open call to NeurIPS-track-adjacent researchers); ratings reflect this group's interpretation of the Croissant RAI rubric. Source-paper sample is biased toward Vision and NLP domains (the two largest fractions of the 102-paper benchmark). Field difficulty is non-uniform: short-text fields (sc:license, sc:datePublished) saw substantially higher pairwise agreement than long-form RAI fields (rai:dataBiases, rai:dataSocialImpact).",
+  "rai:dataLimitations": "Three RAI field definitions (rai:dataCollectionMissingData, rai:dataManipulationProtocol, rai:dataBiases) were tightened in the annotator rubric beyond the canonical Croissant RAI 1.0 specification, which contributes to definitional asymmetry on rai:dataBiases between LLM extractions and human ratings. Annotator demographics are aggregate-only; no per-annotator demographic stratification of agreement is released. Source-paper PDFs are not redistributed; downstream users must obtain them from the original publishers.",
+  "rai:dataUseCases": "Training, Testing, Validation, and Benchmarking of LLM-based metadata extraction systems for Responsible AI documentation. Suitable for evaluating extraction accuracy, abstention calibration (when fields are absent from a paper), and field-specific difficulty profiles.",
+  "rai:isSyntheticData": "False. Annotations are human ratings produced by 22 annotators on real LLM-generated extractions of real published academic papers; no synthetic generation step is part of the annotation pipeline. The 500-paper silver split contains LLM-generated metadata, but the metadata itself is grounded in real source papers rather than synthesised from scratch.",
+  "rai:sourceDatasets": "The 102 gold papers and 500 silver papers are drawn from publicly available academic publications associated with Hugging Face datasets, ranked by download count. Source datasets are listed in the released paper-link manifest (data/paper_links.json) by arXiv identifier; individual paper PDFs are not redistributed and must be obtained from the original publishers (typically arXiv).",
+  "rai:provenanceActivities": "Provenance: (1) Source-paper PDFs were downloaded from arXiv between November 2025 and January 2026 and parsed with PyPDF2 3.0.1. (2) Pre-fill extractions were generated by Claude Sonnet 4.5 (claude-sonnet-4-5-20250929) at temperature 0 with a canonical prompt (SHA-256 prefix 1e1cfdd99246bbf5). (3) Human ratings were collected from 22 annotators between March and April 2026 across seven phases. (4) Gold values were derived by majority vote per (paper, field) cell, with senior-author adjudication for tie-split cells. (5) An end-to-end author audit in May 2026 corrected 191 of 3,060 gold cells (gold_method = 'audit_corrected_2026-05-04'). Each pre-fill extraction's _meta block records the model identifier, prompt SHA-256 prefix, and parser version.",
+  "rai:hasSyntheticData": false,
+  "prov:wasDerivedFrom": [
+    {
+      "@id": "https://huggingface.co/datasets",
+      "prov:label": "HuggingFace Datasets Hub (top-downloaded ML datasets)",
+      "sc:license": "Various (per-dataset)",
+      "prov:wasAttributedTo": {
+        "@id": "https://huggingface.co",
+        "prov:label": "Hugging Face"
+      }
+    },
+    {
+      "@id": "https://arxiv.org",
+      "prov:label": "arXiv",
+      "sc:license": "Per-paper (arXiv non-exclusive license to distribute)",
+      "prov:wasAttributedTo": {
+        "@id": "https://ror.org/05bnh6r87",
+        "prov:label": "arXiv (Cornell University)"
+      }
+    }
+  ],
+  "prov:wasGeneratedBy": [
+    {
+      "@type": "prov:Activity",
+      "prov:type": {
+        "@id": "https://www.wikidata.org/wiki/Q4929239"
+      },
+      "prov:label": "Source paper acquisition",
+      "sc:description": "Source-paper PDFs (102 gold + 500 silver) downloaded from arXiv between November 2025 and January 2026, seeded by the HuggingFace top-downloaded ML datasets list. PDFs parsed with PyPDF2 3.0.1.",
+      "prov:atTime": "2025-11-01T00:00:00Z"
+    },
+    {
+      "@type": "prov:Activity",
+      "prov:type": {
+        "@id": "https://www.wikidata.org/wiki/Q5227332"
+      },
+      "prov:label": "LLM pre-fill extraction",
+      "sc:description": "Claude Sonnet 4.5 (claude-sonnet-4-5-20250929) generated initial metadata extractions at temperature 0 using a canonical prompt (SHA-256 prefix 1e1cfdd99246bbf5). These pre-fills served as the rating substrate for human annotators.",
+      "prov:atTime": "2026-02-15T00:00:00Z"
+    },
+    {
+      "@type": "prov:Activity",
+      "prov:type": {
+        "@id": "https://www.wikidata.org/wiki/Q109719325"
+      },
+      "prov:label": "Human rating + adjudication",
+      "sc:description": "22 annotators rated LLM-generated extractions across 7 phases (March – April 2026). Gold derived by majority rule (>=3 raters) with senior adjudication on disagreements. Final pipeline: 9,595 deduplicated ratings to 3,060 gold cells.",
+      "prov:atTime": "2026-03-01T00:00:00Z"
+    }
+  ]
+}

gold.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b510bf7a81a163a434ac5ec55af0c136c5e97da6c2e0e3807e7b0270bdde8fcd
+size 357473

iaa.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:affcab7689b48213dce2de0a2a2ad686805d2b0f04a51f23f8079afdeea43520
+size 14160

ratings.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0eec928fba8da7df52ab87da403f2448189106252c35222da05d2553c6ed5f1f
+size 560188

silver.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5b77b10bf683e2c05b1a03887f0ab92adebf595df7d78d28dacf8cea30e2f581
+size 37366