hung-k-nguyen's picture
document-classification-benchmark — open-vocab, zero-shot
d204528
|
Raw
History Blame Contribute Delete
3.75 kB
metadata
license: other
license_name: nutrient-eval-restricted
pretty_name: Document Classification Benchmark (open-vocab, zero-shot)
task_categories:
  - zero-shot-image-classification
  - image-classification
language:
  - en
tags:
  - document-ai
  - open-vocabulary
  - zero-shot-image-classification
  - document-image-classification
size_categories:
  - 1K<n<10K
dataset_info:
  features:
    - name: image
      dtype: image
    - name: label
      dtype: string
    - name: candidate_labels
      list: string
    - name: source
      dtype: string
    - name: license_tag
      dtype: string
    - name: doc_id
      dtype: string
    - name: track
      dtype: string
  splits:
    - name: test
      num_bytes: 453775818
      num_examples: 1274
  download_size: 453674933
  dataset_size: 453775818
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/test-*

Document Classification Benchmark (open-vocab, zero-shot)

Given a document image and an arbitrary set of text labels, which one is right? A held-out, zero-shot, open-vocabulary evaluation for document-type classification — labels are supplied at inference, not baked into a head. Test split only; not for training. Every image is drawn from a permissively-licensed, redistributable source.

Powers the document-classification-leaderboard and evaluates document-classification-v2 (commercial) and the open-weight document-classification-v1; try it live in the demo.

Tracks (macro-F1)

  • DocLayNet — document page categories (financial reports, scientific articles, manuals, patents, …).
  • Forms — form vs non-form and fine form sub-types.
  • OOD — document types absent from training (invoices, handwriting, charts, tables) — the open-vocab stress test.
  • OOV — an evaluation protocol (not extra rows): score the same images against never-seen synonym paraphrases of each label ("invoice" → "bill") to measure concept-matching vs memorized wording.

Contents & sources

897 test images (track column selects the slice). All sources are redistributable:

Source Track License
DocLayNet (ds4sd/DocLayNet) doclaynet CDLA-Permissive-1.0
synthetic IRS forms (public-domain templates + faked fields) forms public-domain
mixed permissive HF sources (invoice / handwriting / chart / table) ood per upstream source

Tobacco3482 is deliberately excluded. Its research-only license is not redistributable, so it is not part of this public benchmark (it is used only in internal evaluation).

from datasets import load_dataset
ds = load_dataset("nutrientdocs/document-classification-benchmark", split="test")

Schema

field meaning
image the document image (RGB)
label gold document-type label
candidate_labels the open-vocab candidate set for the row's track
source originating dataset
license_tag redistribution tier of the source
doc_id provenance id
track doclaynet · forms · ood

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.