{
  "@context": {
    "@language": "en",
    "@vocab": "https://schema.org/",
    "arrayShape": "cr:arrayShape",
    "citeAs": "cr:citeAs",
    "column": "cr:column",
    "conformsTo": "dct:conformsTo",
    "containedIn": "cr:containedIn",
    "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",
    "isArray": "cr:isArray",
    "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",
    "rai": "http://mlcommons.org/croissant/RAI/",
    "prov": "http://www.w3.org/ns/prov#"
  },
  "@type": "sc:Dataset",
  "distribution": [
    {
      "@type": "cr:FileObject",
      "@id": "repo",
      "name": "repo",
      "description": "The Hugging Face git repository.",
      "contentUrl": "https://huggingface.co/datasets/anon1285626/MBED/tree/refs%2Fconvert%2Fparquet",
      "encodingFormat": "git+https",
      "sha256": "https://github.com/mlcommons/croissant/issues/80"
    }
  ],
  "recordSet": [],
  "conformsTo": "http://mlcommons.org/croissant/1.1",
  "name": "MBED",
  "description": "anon1285626/MBED dataset hosted on Hugging Face and contributed by the HF Datasets community",
  "alternateName": [
    "anon1285626/MBED"
  ],
  "creator": {
    "@type": "Person",
    "name": "k",
    "url": "https://huggingface.co/anon1285626"
  },
  "keywords": [
    "cc-by-4.0",
    "🇺🇸 Region: US"
  ],
  "license": "https://choosealicense.com/licenses/cc-by-4.0/",
  "url": "https://huggingface.co/datasets/anon1285626/MBED",
  "rai:dataLimitations": "This dataset is confined to the domain of U.S. medical billing and is structurally anchored in American pricing and coding standards (CPT, ICD-10, DRG); it is not applicable to international healthcare systems or private commercial insurance logic that deviates significantly from these norms. The clinical narratives are synthetic and may not capture the full, messy linguistic variance found in real-world physician notes. The benchmark explicitly focuses on eleven perturbation categories; while these cover high-stakes errors like upcoding and quantity inflation, they do not encompass every possible permutation of healthcare fraud or administrative error. This dataset is not designed to evaluate the diagnostic accuracy of models or provide clinical guidance. Statistics derived from it should not be used to draw inferences about any population or disease.",
  "rai:dataBiases": "The seed data is structurally anchored in CMS Medicare claims, meaning medical events and corresponding code-clusters are skewed towards the Medicare-eligible population, specifically older individuals and those with chronic illnesses. Clinical scenarios related to pediatrics, obstetrics, and neonatal care may be under-represented. This benchmark be less effective at flagging model weaknesses in these domains. Furthermore, since the dataset utilizes synthetic narratives to link codes, there is a risk of label bias toward the eleven defined perturbation categories, such as upcoding and phantom billing, which may not capture the full linguistic or fraudulent variance present in private-sector commercial insurance claims. This demographic and structural skew means model behavior should be interpreted as a measure of administrative reasoning within the specific logic of U.S. federal payer systems rather than a universal diagnostic of all healthcare billing types.",
  "rai:personalSensitiveInformation": "This data is free of PII and PHI.",
  "rai:dataUseCases": "The dataset is designed to measure the administrative and financial reasoning capabilities of foundation models within the specific domain of medical billing error detection. It represents the phenomenon of billing inaccuracies, such as quantity inflation and upcoding, which result from the complex alignment of clinical narratives with structured ICD-10 and CPT coding systems. The dataset has been validated as a benchmark for evaluating LLM performance in identifying high-stakes billing perturbations across eleven distinct taxonomic categories. However, this validation does not extend to real-world clinical decision support or the auditing of private commercial insurance claims. Additionally, the dataset is not validated for inferring clinical outcomes or providing medical guidance, as its primary utility is restricted to testing financial audit logic.",
  "rai:dataSocialImpact": "This dataset offers significant positive societal potential by providing a transparent benchmark for identifying medical billing errors that cause financial friction for consumers, though it carries a negative risk of misuse if bad actors attempt to reverse-engineer its logic to generate less detectable fraudulent claims. To mitigate these risks and ensure fairness, the dataset is released as a de-identified, PHI-free resource anchored in public-domain CMS data, intended solely for defensive auditing research rather than clinical or forensic diagnostics. Usage is governed by a focus on administrative reasoning transparency to prevent the dataset from being weaponized against the vulnerable populations it structurally represents.",
  "rai:hasSyntheticData": true,
  "prov:wasDerivedFrom": [
    {
      "@id": "https://www.cms.gov/data-research/statistics-trends-and-reports/medicare-claims-synthetic-public-use-files/cms-2008-2010-data-entrepreneurs-synthetic-public-use-file-de-synpuf/de10-sample-1",
      "prov:label": "CMS DESynPUF Sample 1",
      "sc:license": "Public Domain"
    }
  ],
  "prov:wasGeneratedBy": [
    {
      "@type": "prov:Activity",
      "prov:type": {
        "@id": "https://www.wikidata.org/wiki/Q4929239"
      },
      "prov:label": "clean bill generation",
      "sc:description": "Clean bill generation synthesizes realistic medical bills and after-care summaries from de-identified DE-SynPUF Medicare encounter bundles. The pipeline uses seeded LLM-guided code-cluster generation, geography-constrained hospital selection, and 2025 hospital pricing data with deterministic fallback logic to produce financially coherent clean bill and summary pairs for downstream perturbation and evaluation.",
      "prov:wasAttributedTo": [
        {
          "@type": "prov:SoftwareAgent",
          "@id": "openai:gpt_5_mini",
          "prov:label": "openai:gpt-5-mini"
        }
      ]
    },
    {
      "@type": "prov:Activity",
      "prov:type": {
        "@id": "https://www.wikidata.org/wiki/Q1172378"
      },
      "prov:label": "demographic plausibility filtering",
      "sc:description": "Demographic plausibility filtering reviews generated clean bills for residual patient-facility mismatches and other semantically implausible artifacts that survive earlier synthesis and pricing steps. Flagged cases undergo bounded repair attempts, and unresolved cases are excluded so the released clean set preserves basic clinical and demographic coherence.",
      "prov:wasAttributedTo": [
        {
          "@type": "prov:SoftwareAgent",
          "@id": "openai:gpt_5_mini",
          "prov:label": "openai:gpt-5-mini"
        }
      ]
    },
    {
      "@type": "prov:Activity",
      "prov:type": {
        "@id": "https://www.wikidata.org/wiki/Q4929239"
      },
      "prov:label": "bill perturbation",
      "sc:description": "Bill perturbation injects synthetic billing errors into a seeded subset of clean bills using an LLM-guided mutation pipeline over a predefined taxonomy of medical billing error types. Mutated bills are repriced and normalized so altered services remain financially consistent, while after-care summaries are kept fixed as the clinical ground truth for evaluation.",
      "prov:wasAttributedTo": [
        {
          "@type": "prov:SoftwareAgent",
          "@id": "openai:gpt_5_mini",
          "prov:label": "openai:gpt-5-mini"
        }
      ]
    }
  ]
}