README.md

metadata

license: cc-by-4.0
task_categories:
  - object-detection
language:
  - en
tags:
  - document-ai
  - form-field-detection
  - forms
size_categories:
  - 1K<n<10K
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/test-*

form-field-v1-benchmark

The evaluation set behind the form-field-v1 form-field detectors and the leaderboard — 5,914 form pages with 126,938 annotated fields (Text, Choice = checkbox/radio, Signature), across three variants:

variant	images	fields	notes
empty	2,334	49,171	blank forms
filled	1,826	39,797	digitally-typed values
handwritten	1,754	37,970	handwritten (ink) values

A Nutrient dataset, released under CC-BY-4.0.

• 🎯 Demo · 🏆 Leaderboard

Columns

• image — the rendered page (RGB).
• page_id — stable id; ends in _empty / _filled / _handwritten.
• variant — empty | filled | handwritten.
• width, height — page pixel size.
• fields — list of { "box": [x, y, w, h], "category": "Text"|"ChoiceButton"|"Signature" } (page pixels).
• source, license_class — provenance tags.

Scoring

Detectors are scored with COCO mAP50-95 (pycocotools, maxDets [1,100,1000]), reported overall and per variant.

# pip install datasets pycocotools
python score_detector.py your_predictions.json
# predictions: [{"image_id": page_id, "category_id": 1|2|3, "score": s, "bbox": [x,y,w,h]}]

Reference numbers (mAP50-95) on this set — the form-field-v1 family: medium 0.529 / 0.783 / 0.764, small 0.456 / 0.716 / 0.680, nano 0.353 / 0.560 / 0.526 (empty / filled / handwritten).

License

CC-BY-4.0. Attribute to Nutrient. Empty pages are real blank forms; filled and handwritten values are synthetic. No PII.

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.