{
  "@context": {
    "@language": "en",
    "cr": "http://mlcommons.org/croissant/",
    "dct": "http://purl.org/dc/terms/",
    "rai": "http://mlcommons.org/croissant/RAI/",
    "sc": "https://schema.org/"
  },
  "@type": "sc:Dataset",
  "name": "ERP-Bench",
  "alternateName": "anonymouse321/erp-bench",
  "url": "https://huggingface.co/datasets/anonymouse321/erp-bench",
  "dct:conformsTo": "http://mlcommons.org/croissant/RAI/1.0",
  "rai:dataCollection": "ERP-Bench is generated by the Anchor pipeline from parametric procurement and manufacturing specifications. Each accepted task is sampled from configured ranges, solved with OR-Tools CP-SAT, and compiled into aligned natural-language instructions, Odoo environment setup files, oracle solution scripts, and terminal-state verifiers.",
  "rai:dataCollectionType": [
    "Software Collection",
    "Experiments",
    "Others: synthetic scenario generation"
  ],
  "rai:dataCollectionRawData": "The dataset does not collect user records or operational business records. Scenario entities, company names, product names, customer orders, vendors, prices, deadlines, inventory quantities, BOMs, workcenter capacities, and invoices are synthetic artifacts generated by the benchmark code.",
  "rai:dataCollectionMissingData": "Not applicable. Tasks are generated from structured specifications; samples that are infeasible or fail solver certification are rejected before release.",
  "rai:dataPreprocessingProtocol": [
    "Sample parameter settings from checked-in procurement dataset configs.",
    "Reject infeasible samples through pre-solver checks and CP-SAT certification.",
    "Render accepted samples into Harbor task directories containing instructions, Odoo setup scripts, oracle solutions, and verifier code."
  ],
  "rai:dataImputationProtocol": "No missing source records are imputed. The generator rejects invalid or infeasible samples rather than filling missing observational data.",
  "rai:dataManipulationProtocol": "Dataset items are produced by deterministic template rendering from solver-certified scenario builds. No private source data is transformed or anonymized.",
  "rai:dataAnnotationProtocol": "No human annotation labels are packaged with the dataset. Task correctness labels and optimal plans are produced by the CP-SAT solver and deterministic rendering pipeline.",
  "rai:dataAnnotationAnalysis": "Not applicable. There is no aggregation of human annotations or adjudication of annotator disagreement.",
  "rai:annotationsPerItem": "No human labels per item.",
  "rai:machineAnnotationTools": [
    "OR-Tools CP-SAT",
    "ERP-Bench procurement generator",
    "Jinja2 task artifact templates"
  ],
  "rai:personalSensitiveInformation": [
    "No intentional personal or sensitive information is included.",
    "Business entities and names are synthetic and should be treated as fictional.",
    "Generated tasks include local Odoo runtime credentials and API-key file paths for isolated benchmark containers only; these are not real external credentials."
  ],
  "rai:dataUseCases": [
    "Evaluation of AI agents on long-horizon ERP procurement and manufacturing workflows.",
    "Comparison of coding, browser, and computer-use agent scaffolds using the same terminal-state verifier.",
    "Research on artifact-drift-resistant benchmark generation, constrained business planning, and agent reliability.",
    "Development or regression testing of agents intended to operate in sandboxed Odoo-like business software environments."
  ],
  "rai:dataLimitations": [
    "ERP-Bench is synthetic and does not represent production customer, supplier, employee, or financial records.",
    "The benchmark covers procurement and manufacturing workflows in Odoo 19; results may not generalize to all ERP modules, companies, localizations, or business policies.",
    "The verifier grades terminal Odoo state for encoded constraints, optimality, and traceability; it does not judge every possible business preference or UI interaction quality.",
    "Agents may overfit to generated task structure, Odoo version details, or benchmark-specific verifier behavior.",
    "The dataset is intended for sandboxed evaluation and should not be used as operational business guidance."
  ],
  "rai:dataBiases": [
    "Scenario distributions reflect the authored generator, product domains, constraint families, and difficulty configurations rather than a statistically representative sample of real ERP work.",
    "Synthetic names, industries, products, prices, and deadlines may encode generator design choices and may not reflect geographic, economic, or organizational diversity.",
    "The benchmark emphasizes planning and execution tasks that are amenable to CP-SAT formalization, which may underrepresent qualitative, interpersonal, or exception-heavy business workflows."
  ],
  "rai:dataSocialImpact": "ERP-Bench can support more transparent evaluation of agents in business software, but benchmark performance should not be interpreted as readiness for live procurement or manufacturing operations without separate security, reliability, compliance, and human-oversight review.",
  "rai:dataReleaseMaintenancePlan": "Releases are regenerated from checked-in code, configs, seeds, schemas, and templates. Known task defects should be fixed at the generator or template level and propagated by deterministic regeneration. Future releases should preserve compatibility notes for Odoo version, task format, and generation configs."
}