Datasets:

nutrientdocs
/

wholedoc-bench-sample

license: other
pretty_name: WholeDoc Bench (sample)
task_categories:
  - image-to-text
language:
  - de
  - en
  - es
  - fr
  - nl
size_categories:
  - n<1K
tags:
  - document-parsing
  - pdf
  - benchmark
  - evaluation
  - tables
  - charts
  - ocr
  - layout-detection
  - barcode
  - mrz
configs:
  - config_name: default
    data_files:
      - split: test
        path: documents.jsonl
citation: |
  @misc{wholedocbench2026,
    title  = {WholeDoc Bench: Continuity, Fidelity and Machine-Readable Content in Document Parsing},
    author = {Moldovan, Denis and Sookun, Shrevin and Mogage, Nicolae and Pizur, Matus and Mitrea, Costinel and Houssin, Olivier and Alby, Etienne and Carlier, Coralie and Mart{\'\i}n, Daniel},
    year   = {2026},
    url    = {https://github.com/PSPDFKit-labs/wholedoc-bench}
  }

WholeDoc Bench

Quick links: [🌐 Leaderboard] [📜 Paper] [💻 Code]

This is the public sample of WholeDoc Bench, a benchmark for evaluating document parsers on whole documents. It contains 10 documents (28 pages) with their ground-truth answer keys, so you can inspect the annotation format and run the evaluation pipeline on these documents.

Results from the full private evaluation set are available on the benchmark page and in the paper. Those results cannot be reproduced from this sample alone. The evaluation code is available to run and score supported parsers.

Usage

Download the whole release, pinned to its tag. The rows alone are not enough to score:

uvx --from huggingface_hub hf download nutrientdocs/wholedoc-bench-sample --repo-type dataset --revision v1 --local-dir wholedoc-bench-sample

Then run a parser, score it and build a leaderboard with the evaluation code:

uv run wholedoc-bench run --provider pdftotext --corpus wholedoc-bench-sample --out runs/pdftotext
uv run wholedoc-bench score --corpus wholedoc-bench-sample --predictions runs/pdftotext --results results
uv run wholedoc-bench report --results results

Every command checks the release first and stops if a file is missing or changed.

Files and row fields

documents.jsonl       one row per document
documents/<id>.pdf    the documents
groundtruth/<id>.json the ground truth of each document
release.json          the release tag, its source commit and counts
SHA256SUMS            the checksum of every file above
published.jsonl       every document id ever published, with its first release

Field	Meaning
`id`	the document's permanent id
`document_path`, `document_sha256`	the document file and its checksum
`format`, `pages`	file format and page count
`language`, `scan`	languages, and whether the document is a scan
`tracks`, `supported_metrics`	what this document is scored on
`groundtruth_source`	how the ground truth was made
`groundtruth_path`, `groundtruth_sha256`	the ground truth file and its checksum
`category`, `limits`, `licence`, `attribution`	not recorded in this release

Public sample and private set

The published benchmark results use the full private evaluation set. This public sample lets you inspect the answer keys and run the evaluation pipeline, but it does not reproduce the full benchmark results.

Citation

@misc{wholedocbench2026,
  title  = {WholeDoc Bench: Continuity, Fidelity and Machine-Readable Content in Document Parsing},
  author = {Moldovan, Denis and Sookun, Shrevin and Mogage, Nicolae and Pizur, Matus and Mitrea, Costinel and Houssin, Olivier and Alby, Etienne and Carlier, Coralie and Mart{\'\i}n, Daniel},
  year   = {2026},
  url    = {https://github.com/PSPDFKit-labs/wholedoc-bench}
}

About the author

This project is maintained and funded by Nutrient - The deterministic document infrastructure enterprises run their highest-stakes workflows on: replayable output, clear exceptions, and full audit trails on the messy, regulated documents where AI alone breaks.