Datasets:

nutrientdocs
/

form-field-vlm-benchmark

license: cc-by-4.0
pretty_name: Form Field Detection Benchmark
task_categories:
  - object-detection
  - image-to-text
language:
  - en
tags:
  - document-ai
  - forms
  - form-understanding
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/test-*

Form Field Detection Benchmark

A fixed, test-only benchmark for end-to-end form-field extraction. It contains the exact 100 clean form pages and 701 fields used by the form-field-vlm scorecard and leaderboard. Try the complete server-side pipeline in the demo.

The task is to return every interactive widget as constrained JSON with:

box: [x0, y0, x1, y1] on a 0–1000 page grid
type: text, choice_checkbox, choice_radio, choice_select, or signature
label: the field's text label
group_id: shared radio-group identifier, otherwise null

Metrics

The headline is typed-field F1 at IoU 0.5. A match requires sufficient box overlap and the correct fine field type. We also report IoU 0.2, per-type precision/recall/F1, Box F1, Box Recall, density slices, count MAE, false positives per page, and measured seconds per page.

IoU 0.5 is the ranking metric because downstream form filling needs tight widget boxes. IoU 0.2 is a diagnostic for approximately correct but loosely grounded predictions.

Schema

from datasets import load_dataset
benchmark = load_dataset("nutrientdocs/form-field-vlm-benchmark", split="test")

Field	Meaning
`image`	RGB form-page image
`page_id`	Stable benchmark page identifier
`width`, `height`	Original image dimensions
`fields`	Gold field objects in original page coordinates
`source`	Upstream source family
`license_class`	Redistribution classification

Scoring

The reference scorer accepts a JSON list of predictions in original page pixels:

python score.py --benchmark-repo nutrientdocs/form-field-vlm-benchmark \
  --predictions predictions.json --out result.json

Each prediction is {page_id, box: [x,y,w,h], type, label?, group_id?} in original page pixels. The self-contained scorer downloads the public test split and emits the complete IoU 0.5 and 0.2 breakdown used by the leaderboard. Submit the resulting JSON with the model ID, license, runtime environment, and seconds/page for review.

Scope

This fixed 100-page set is small enough to run across local and cloud VLMs, but it should not be treated as a comprehensive measure of all form styles. A future, larger document-disjoint benchmark should be published separately rather than silently changing this set.

License and attribution

The benchmark pages derive from CommonForms, released under CC-BY-4.0. See ATTRIBUTION.md for the citation and provenance statement.

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.