DocuBench / docs /submissions.md
urimer's picture
Initial upload: 50 documents, schemas, hand-verified labels, scorer, baseline results
ca66b51 verified
|
Raw
History Blame Contribute Delete
2.22 kB

Submission Format

DocuBench result sets live under:

results/<system_name>/<doc_id>.json

<system_name> should be lowercase, filesystem-safe, and stable across reports.

Required Per-Document File

Each document result should be a JSON object with a top-level data key:

{
  "data": {
    "invoice_number": "INV-001",
    "line_items": []
  }
}

The data object is scored against labels/<doc_id>.json using schemas/<doc_id>.json.

Recommended Metadata

Include run metadata when available:

{
  "data": {},
  "cost": 0.0,
  "time_sec": 0.0,
  "meta": {
    "system": "example-system",
    "model": "model-or-api-version",
    "run_date": "2026-06-26",
    "configuration": "short human-readable config"
  }
}

If a system fails to complete, refuses, or returns output that cannot be parsed into the requested schema, still write a result file with an empty data object:

{
  "status": "failed",
  "error": {
    "type": "schema_mismatch",
    "message": "model output did not conform to the schema"
  },
  "data": {},
  "meta": {
    "model": "model-or-api-version"
  }
}

This makes the failure visible and causes all non-empty labeled fields for that document to be counted as errors.

Validation And Scoring

Run:

docubench validate
docubench score --engine <system_name>

To regenerate reports for all result sets:

docubench report

Prompts And Run Configuration

Commit the prompt and run configuration that produced a submission under prompts/, following the existing baselines. The committed LLM baselines load their instruction from prompts/extraction_prompt.txt at runtime, and each result file stamps its model id, provider, and schema_mode in meta — together these make a result set reproducible and auditable.

Review Expectations

Public submissions should state:

  • system name and version
  • model or provider version, if applicable
  • extraction settings that affect output
  • date of run
  • whether the run was produced by released code, proprietary code, or manual workflow
  • any documents skipped or failed
  • cost and latency if measured