metadata.json

{
  "@context": {
    "@language": "en",
    "@vocab": "https://schema.org/",
    "citeAs": "cr:citeAs",
    "column": "cr:column",
    "conformsTo": "dct:conformsTo",
    "cr": "http://mlcommons.org/croissant/",
    "data": {"@id": "cr:data", "@type": "@json"},
    "dataBiases": "cr:dataBiases",
    "dataCollection": "cr:dataCollection",
    "dataType": {"@id": "cr:dataType", "@type": "@vocab"},
    "dct": "http://purl.org/dc/terms/",
    "extract": "cr:extract",
    "field": "cr:field",
    "fileProperty": "cr:fileProperty",
    "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",
    "personalSensitiveInformation": "cr:personalSensitiveInformation",
    "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",
  "name": "FinCriticalED",
  "conformsTo": "http://mlcommons.org/croissant/1.0",
  "citeAs": "He, Y. et al. (2026). FinCriticalED: A Financial Document OCR Evaluation Benchmark for Critical Entity Detection. ACM MM 2026.",

  "description": "FinCriticalED is a benchmark dataset for evaluating Large Language Models on their ability to identify financially critical values in complex financial documents, such as profit and revenue figures, important legal compliance due dates, and contract durations. The dataset contains 859 document pages from real-world financial documents (2023–2026) including 10-K/10-Q reports, Form 4 and Form 8-K filings, Form 990 tax records, securities transaction records, and legal agreements (M&A documents, credit agreements, bond indentures). It consists of two parts: (1) raw_put — base64-encoded page images paired with their matched HTML source chunks; and (2) gold_annotation_html — expert-annotated HTML files with inline entity tags marking five types of financially critical facts (9,481 total annotations, Fleiss κ = 0.8837).",

  "url": "https://huggingface.co/datasets/TheFinAI/FinCriticalED",
  "keywords": [
    "OCR",
    "financial documents",
    "information extraction",
    "named entity recognition",
    "benchmark",
    "SEC filings",
    "vision-language models",
    "document understanding",
    "financial OCR",
    "key information extraction"
  ],
  "license": "https://www.apache.org/licenses/LICENSE-2.0",
  "creator": [
    {"@type": "sc:Person", "name": "Yueru He", "email": "yh3507@columbia.edu"},
    {"@type": "sc:Person", "name": "Xueqing Peng"},
    {"@type": "sc:Person", "name": "Yan Wang"},
    {"@type": "sc:Person", "name": "Yi Han"},
    {"@type": "sc:Organization", "name": "The FinAI Team"}
  ],
  "datePublished": "2026",
  "inLanguage": "en",
  "isAccessibleForFree": true,
  "isLiveDataset": false,

  "distribution": [
    {
      "@type": "cr:FileSet",
      "@id": "raw-put-files",
      "name": "raw_put",
      "description": "Raw input files. Each JSON file contains one document page sample with two fields: 'image' (base64-encoded PNG of the rendered financial document page) and 'matched_html' (the corresponding HTML chunk from the source SEC filing that matches the page image).",
      "containedIn": {"@id": "huggingface-repo"},
      "encodingFormat": "application/json",
      "includes": "data/raw_put/*.json"
    },
    {
      "@type": "cr:FileSet",
      "@id": "gold-annotation-html",
      "name": "gold_annotation_html",
      "description": "Annotated gold standard HTML files. Each .txt file (gold_N.txt) contains one post-processed HTML document for the corresponding sample N in raw_put. Five custom inline tags mark financially critical facts: <number> (numeric values), <temporal> (dates and durations), <monetaryunit> (currency symbols and scale indicators), <reportingentity> (named organizational/personal entities), <financialconcept> (domain-specific financial terms).",
      "containedIn": {"@id": "huggingface-repo"},
      "encodingFormat": "text/html",
      "includes": "data/gold_annotation_html/gold_*.txt"
    },
    {
      "@type": "cr:FileObject",
      "@id": "huggingface-repo",
      "name": "TheFinAI/FinCriticalED HuggingFace repository",
      "description": "HuggingFace dataset repository containing all raw input and annotated gold standard files.",
      "contentUrl": "https://huggingface.co/datasets/TheFinAI/FinCriticalED",
      "encodingFormat": "git+https"
    }
  ],

  "recordSet": [
    {
      "@type": "cr:RecordSet",
      "@id": "raw-input-records",
      "name": "Raw Input Records",
      "description": "One record per document page. Each record contains a base64-encoded page image and its matched HTML source chunk. Sample N corresponds to gold_N.txt in the annotation files.",
      "field": [
        {
          "@type": "cr:Field",
          "@id": "raw-input-records/sample_id",
          "name": "sample_id",
          "description": "Zero-indexed sample identifier (N in raw_put/N.json and gold_annotation_html/gold_N.txt).",
          "dataType": "sc:Integer"
        },
        {
          "@type": "cr:Field",
          "@id": "raw-input-records/image",
          "name": "image",
          "description": "Base64-encoded PNG image of the rendered financial document page (derived from intermediate PDF rendering of the SEC EDGAR or Candid.org HTML source).",
          "dataType": "sc:ImageObject",
          "source": {
            "fileSet": {"@id": "raw-put-files"},
            "extract": {"jsonPath": "$.image"}
          }
        },
        {
          "@type": "cr:Field",
          "@id": "raw-input-records/matched_html",
          "name": "matched_html",
          "description": "HTML chunk from the original SEC filing that corresponds to the page image. Used as the ground truth source for annotation.",
          "dataType": "sc:Text",
          "source": {
            "fileSet": {"@id": "raw-put-files"},
            "extract": {"jsonPath": "$.matched_html"}
          }
        }
      ]
    },
    {
      "@type": "cr:RecordSet",
      "@id": "gold-annotation-records",
      "name": "Gold Annotation Records",
      "description": "One record per document page (gold_N.txt). Contains the post-processed HTML with inline entity tags marking all financially critical facts. The suffix N matches the sample_id in raw-input-records.",
      "field": [
        {
          "@type": "cr:Field",
          "@id": "gold-annotation-records/sample_id",
          "name": "sample_id",
          "description": "Zero-indexed sample identifier matching the corresponding raw input record.",
          "dataType": "sc:Integer",
          "references": {"field": {"@id": "raw-input-records/sample_id"}}
        },
        {
          "@type": "cr:Field",
          "@id": "gold-annotation-records/annotated_html",
          "name": "annotated_html",
          "description": "Full HTML of the document page with five custom entity tags: <number>, <temporal>, <monetaryunit>, <reportingentity>, <financialconcept>. Surrounding HTML structure is preserved verbatim from the source filing.",
          "dataType": "sc:Text",
          "source": {
            "fileSet": {"@id": "gold-annotation-html"},
            "extract": {"fileProperty": "fulltext"}
          }
        }
      ]
    }
  ],

  "dataCollection": "Source data covers real-world financial documents from 2023–2026, including tabular and narrative-heavy formats: 10-K and 10-Q reports, Form 4 and Form 8-K filings, Form 990 tax records, securities transaction records, and legal agreements (M&A documents, credit agreements, bond indentures). HTML documents were downloaded from SEC EDGAR (https://www.sec.gov/search-filings) and tax forms from Candid.org (https://app.candid.org/, Data for Academics program; attribution: Powered by Candid, © Candid). Documents were converted to page images via intermediate PDF rendering. The annotation pipeline proceeded in four phases: (1) bounding-box layout annotation in Label Studio; (2) OCR extraction per region via Tesseract; (3) QA and reasoning generation via GPT-4o; (4) entity-level span annotation by human experts with multi-round calibration against a detailed annotation guideline. Annotators were provided rendered HTML chunks in Label Studio alongside corresponding page images for reference.",

  "dataBiases": "The dataset is limited to English-language financial documents and may underrepresent other languages and regions. Source documents are drawn from U.S. financial reporting contexts (SEC EDGAR, Candid.org). Form 4 insider transaction filings are over-represented in the SEC Required Filings category. The benchmark emphasizes numeric and temporal entities; other critical financial information such as company names and person names is present but less emphasized. Documents are rendered as digital-native images (300 dpi via Chromium) and do not include scanned or physically degraded documents, which may limit generalization to real-world scan noise.",

  "personalSensitiveInformation": "The FinCriticalED dataset does not contain any personally identifiable information (PII) and is strictly focused on English-language open-source financial documents. All source documents are publicly available financial records. No personal or sensitive information beyond what is already publicly disclosed in SEC filings and Form 990 tax records is present in this dataset."
}