You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
By accessing this dataset you agree to the SymageDocs Synthetic Dataset License v1.0 (see the LICENSE file). You may train, fine-tune, and evaluate models on the data for your own internal use, including commercially. You may NOT redistribute the dataset or use it to build a competing dataset.
Log in or Sign Up to review the conditions and access this dataset content.
SymageDocs — Coherent US Tax / Health / Employment Forms (FUNSD)
A fully synthetic document-AI training set: three US forms — IRS Form 1040, CMS-1500, and USCIS Form I-9 — filled from the same synthetic identity, so name / SSN / address / employer flow consistently across all three renderings. Each page ships with FUNSD ground truth (word boxes, entity labels, key–value linking) plus a LayoutLMv3-ready token/bbox/tag view.
- 3,000 page-level image + annotation rows (train 2,400 / test 600)
- 3 correlated forms per identity, the configured number of coherent identities
- Clean typed renderings; splits drawn at the identity level (no package leaks across train/test)
Why this is different
Public form datasets (FUNSD, CORD, DocLayNet) give isolated pages. This set gives coherent multi-form identities: the same person's 1040, CMS-1500, and I-9, with field-level ground truth linked across documents — closer to a real onboarding/claims packet than a bag of unrelated scans.
Dataset structure
| Field | Type | Description |
|---|---|---|
image |
image | Clean typed page render (PNG) |
tokens |
list[str] | Word tokens (FUNSD word order) |
bboxes |
list[[x0,y0,x1,y1]] | Per-token boxes, normalized 0–1000 (LayoutLM) |
ner_tags |
ClassLabel seq | BIO tags: O, B-HEADER, I-HEADER, B-QUESTION, I-QUESTION, B-ANSWER, I-ANSWER |
funsd_json |
str | Full FUNSD form array — entities, boxes, key–value linking, checkboxes |
form_id |
str | Which form this page belongs to |
identity_id |
int | Coherence key — join a person's pages across forms |
page |
int | Page index within the form |
Forms:
| form_id | Name | Pages |
|---|---|---|
irs_f1040_modern_2024 |
Form 1040 - U.S. Individual Income Tax Return | 2 |
cms_1500_standard_02_12 |
CMS-1500 Health Insurance Claim Form | 1 |
i9_standard_2024 |
I-9 Employment Eligibility Verification | 4 |
Blank pages are intentional
I-9 Supplement A (Preparer/Translator) and Supplement B (Reverification/Rehire)
are optional pages most employees never trigger, so the large majority render
blank — exactly as they sit in a real HR file. A portion of the rows are
therefore legitimately empty supplement pages (a page image + empty field boxes,
zero tokens). This is deliberate realism, not missing data: a model that reads
real I-9s must also recognize a blank supplement. For a pure token-classification
subset, filter to rows where num_entities > 0 (or len(tokens) > 0).
Load it
from datasets import load_dataset
ds = load_dataset("Symage/us-tax-health-employment-forms")
print(ds["train"][0]["tokens"][:10], ds["train"][0]["ner_tags"][:10])
What else SymageDocs can generate
This dataset is one narrow slice — three forms, clean typed renders, FUNSD labels. The SymageDocs engine that produced it can emit far more, on demand:
- Annotation formats: FUNSD (here), BIO token NER, YOLO & COCO field-region detection, Donut image→JSON imagefolder, plus raw per-field ground-truth JSON and CSV.
- Document renders: typed PDF, handwritten PDF, pre-filled PDF, typed & handwritten PNG, at configurable DPI.
- Realism / degradation: clean (here) through scanned & noised profiles (skew, blur, JPEG, ink bleed, coffee stains, …) at graded intensity, plus configurable ink color.
- Coverage: 65+ form types across tax, healthcare, insurance, HR, onboarding, financial & commercial categories — any subset as a coherent multi-form identity package, at arbitrary volume.
Need other forms, formats, degradation profiles, or a larger correlated corpus? That's the paid product → https://symagedocs.ai
Ethical considerations & synthetic-data notice
Every record is algorithmically generated and fully synthetic — not derived from any real person's records. Because values are generated programmatically, the data may coincidentally resemble real names/addresses; it is not real PII. The form layouts are US-government public domain (IRS 1040, USCIS I-9) or NUCC public domain (CMS-1500); medical procedure codes are HCPCS Level II / synthetic (no licensed AMA CPT descriptor text).
License
Gated under the SymageDocs Synthetic Dataset License v1.0 (LICENSE).
Evaluation + commercial ML training for your own internal use are permitted;
redistribution and building a competing dataset are not. Access is gated with
click-through acceptance (auto-approved).
Citation
@misc{symagedocs_forms,
title = {SymageDocs — Coherent US Tax / Health / Employment Forms (FUNSD)},
author = {Symage, Inc.},
year = {2026},
url = {https://symagedocs.ai}
}
- Downloads last month
- 11