Symage/synthetic-us-forms-preview

SymageDocs — Synthetic US Forms Preview

A small, CC-BY-4.0, fully synthetic document-AI training set: 525 labeled page images across six families of US business and government forms, each page shipping FUNSD ground truth plus a LayoutLM-ready token/bbox/tag view.

This is a preview subset. It exists so you can load real output from the SymageDocs generator, inspect the label quality, and decide whether generating your own corpus is worth your time — without an account, an email, or a click-through.

• 525 page-level rows (train 420 / test 105)
• 10 form templates across 6 families
• Clean typed renders at 300 DPI
• CC-BY-4.0 — use it commercially, redistribute it, remix it. Just credit SymageDocs.

What's in it

form_id	Form	Documents	Pages each
irs_w2_single_page_2026	IRS Form W-2 (single page, 2026)	150	1
irs_f1040_modern_2024	IRS Form 1040 (2024)	50	2
cms_1500_standard_02_12	CMS-1500 health insurance claim (02/12)	50	1
irs_w9_standard_2024	IRS Form W-9 (2024)	50	1
i9_standard_2024	USCIS Form I-9 (2024)	25	up to 3
invoice_classic … invoice_construction	Commercial invoices, 5 layouts	20 each	1

Form layouts are US-government public domain (IRS, USCIS) or NUCC public domain (CMS-1500); the invoice layouts are original to SymageDocs. Medical procedure codes are HCPCS Level II or synthetic — no licensed AMA CPT descriptor text appears anywhere in this data.

Schema

Column	Type	Description
image	image	Clean typed page render, 300 DPI PNG
tokens	list[str]	Word tokens in FUNSD reading order
bboxes	list[[x0,y0,x1,y1]]	Per-token boxes, normalized 0–1000 (LayoutLM convention)
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, checkbox states
coco_json	str	COCO-style field-region detection annotations for the page — one box per form field
yolo_txt	str	YOLO-format per-word boxes for the page — one line per rendered word, classed by structural role
form_id	str	Which form this page belongs to
identity_id	int	Document key — groups a multi-page form's pages
page	int	Page index within the document
num_entities	int	Entity count (see the note on unfilled supplement pages)

Four label views ship for every page — FUNSD, BIO token tags, COCO, and YOLO — so you can benchmark token classification and field-region detection off the same images without re-annotating.

coco_json and yolo_txt are different granularities, on purpose. COCO boxes are field regions (one per form field); YOLO boxes are words (one per rendered word). They are complementary views, not the same annotation in two file formats — do not train a detector on one and evaluate it against the other. For field-region detection use coco_json; for word-level detection or reading order use yolo_txt (or bboxes, which is the same granularity in LayoutLM convention).

yolo_txt class ids are a fixed 6-class structural vocabulary, constant across every SymageDocs dataset and aligned with the FUNSD taxonomy:

0 question
1 answer
2 header
3 label
4 instruction
5 other

Splits are drawn at the document level, so a multi-page form never straddles train and test.

Unfilled supplement pages are intentional

I-9 Supplement A (Preparer/Translator) and Supplement B (Reverification/Rehire) are optional pages most employees never trigger, so they render with no values filled in — exactly as they sit in a real HR file. Those rows still carry the page image and the supplement's labeled-but-empty field regions (funsd_json, coco_json and yolo_txt are all populated; num_entities is ~12) but zero tokens. A model that reads real I-9s must also recognize an unfilled supplement.

For a pure token-classification subset, filter on len(tokens) > 0. num_entities > 0 does not exclude these rows.

Load it

from datasets import load_dataset

ds = load_dataset("Symage/synthetic-us-forms-preview")
row = ds["train"][0]
print(row["form_id"], row["tokens"][:10], row["ner_tags"][:10])
row["image"].show()

The tokens / bboxes / ner_tags triple is already in the shape LayoutLMv3Processor expects, so the standard funsd-layoutlmv3 fine-tuning recipe runs against this set unchanged.

How it was generated

Every record is produced by the SymageDocs generation engine: a synthetic identity is sampled, propagated through a form's field bindings and computed fields, rendered to PDF, and rasterized — with the word-level annotations captured from the renderer itself rather than recovered by OCR. The boxes are exact by construction, not estimated, and no OCR error is baked into the ground truth.

Generation runs through the same production code path that serves paying customers, from a pinned seed, so this subset is reproducible rather than a hand-curated showcase.

Zero PII by construction

There is no real personal data here, because there was never any real personal data in the pipeline: names, addresses, SSNs, EINs, NPIs, and account numbers are all generated from synthetic distributions and validated check-digit schemes. Nothing was scraped, de-identified, or anonymized from real records — so there is no re-identification risk to reason about.

Because values are generated programmatically, a record may coincidentally resemble a real person's details. It is still synthetic, and it is not derived from anyone's records.

Honest limitations

Worth knowing before you build on it:

• Clean renders only. No scanner noise, skew, blur, or ink bleed in this subset. Models trained on it alone will not be robust to photographed or faxed documents. The generator produces graded degradation profiles; this preview deliberately ships the clean tier.
• Small. It is a preview, sized to be inspected, not to saturate a model.
• Independent documents. Each document has its own identity. Cross-document coherence is a separate dataset (below).
• US forms, English only.
• No published benchmark. We have not trained a reference model on this subset, so we quote no accuracy numbers. Judge it by inspecting the labels.

Related: the full coherent dataset

Symage/coherent-forms-1040-cms1500-i9 is a larger set where one synthetic person's IRS 1040, CMS-1500, and I-9 are all filled from the same identity — name, SSN, address, and employer flow consistently across all three documents, which is much closer to a real onboarding or claims packet than a bag of unrelated pages. It is free, behind a one-click license acceptance, under a different (non-CC-BY) license.

Coherent multi-form packages are the thing the generator does that a static dataset can't hand you; this preview is the single-document taste of it.

Generate your own

This subset is a fixed slice. The engine behind it takes parameters:

• 50+ form types — tax, healthcare, insurance, HR and onboarding, financial, commercial. Browse the catalog at symagedocs.ai/forms.
• Label formats — FUNSD, BIO, COCO, YOLO, Donut image→JSON, plus raw per-field ground truth as JSON and CSV.
• Renders — typed and handwritten PDF, pre-filled PDF, PNG at your DPI.
• Degradation — clean through heavily scanned, graded intensity: skew, blur, JPEG artifacts, ink bleed, stains.
• Coherent multi-form identity packages at whatever volume you need.

pip install symagedocs

The free tier starts you with 1,000 credits (a promotional 500/month allowance plus a 500-credit welcome bonus) — enough to generate a real corpus and check it against your pipeline before deciding anything. → symagedocs.ai

License

CC-BY-4.0. You may use, modify, and redistribute this subset, including commercially and including for model training. Please credit SymageDocs and link back to symagedocs.ai.

Note that the related coherent dataset linked above is under different, more restrictive terms — this CC-BY-4.0 grant covers this preview subset only.

Citation

@misc{symagedocs_us_forms_preview,
  title  = {SymageDocs — Synthetic US Forms Preview (FUNSD / LayoutLM)},
  author = {Symage, Inc.},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Symage/synthetic-us-forms-preview}
}