caiotheodoro/nfse-synth

NFS-e Synth

Synthetic Brazilian service invoices (NFS-e, Nota Fiscal de Serviço eletrônica) for field extraction and arithmetic validation. There are 10,000 page images in four layout families. Each image has exact gold fields, and 20% carry known injected violations that a deterministic checker recovers exactly.

Everything here is synthetic. No real invoice, person, company or municipal form is included. Each page is watermarked SINTÉTICA — SEM VALOR FISCAL. The layout families copy the structure of four real municipal forms: which fields are printed, where they sit, and how ISS is rounded. They carry no logos, coats of arms, or real identifiers. CPF/CNPJ numbers are randomly generated with valid check digits, so any match with a real document is coincidental.

Why it exists

No public NFS-e extraction benchmark exists. The real notes that motivated this set are accounting-client data covered by LGPD art. 11 and cannot be released. The set pairs with a private, pre-registered study on 604 real units. That study reports aggregate results only (see Related work).

Structure

data/{train,validation,test}/ holds the images plus a metadata.jsonl per split (Hugging Face imagefolder).

split	notes
train	8,000
validation	1,000
test	1,000

Each of the four families contributes 2,500 notes. The split is by seed within each family.

family	modelled on	ISS rounding	printed fields	scan artefacts
bc_like	Balneário Camboriú	half-up	every field, including federal rates	no (PNG)
sp_like	São Paulo	truncate	docs, gross, deductions, base, rate, ISS, withholding amounts	yes (JPEG)
barueri_like	Barueri	half-up	docs, items, federal withholdings, net ("fatura")	no (PNG)
osasco_like	Osasco	half-up	docs, gross, base, rate, ISS, IR	yes (JPEG)

sp_like truncates ISS because São Paulo does. In the real notes, 22 of the 24 cases where truncation and half-up rounding disagree match truncation, and none match half-up.

Row fields

• file_name, id, family, seed
• iss_rounding: truncate or half_up
• visible_fields: the fields this family prints. Score only these. The remaining gold fields are hidden truth, kept so the checker can verify the note.
• gold: the full note (CPF/CNPJ as digits, money and rates as decimal strings, items list)
• violations: the oracle rules this note breaks (empty for clean notes)
• has_violation, descriptions, parties (synthetic names/addresses), number, date

The checker (oracle)

Every note has to satisfy these rules unless a violation was injected:

1. federal withholding = base × rate (PIS, COFINS, INSS, IR, CSLL)
2. ISS = base × rate, with the family's rounding
3. base = gross − deductions − unconditional discount
4. net = gross − withholdings − discounts − ISS if withheld
5. items sum to gross; each item total = qty × unit price
6. valid CPF/CNPJ check digits

Injected violations only target rules that can be checked from the family's printed fields. On every row, the published gold reproduces exactly the published violations under the family's rounding. This is enforced at build time and in tests/test_dataset.py.

Suggested tasks

• Extraction: image → the visible_fields. Score by field exact match after numeric normalization.
• Validation: image → does this note break any rule, and which one? Labels are exact.
• Self-consistency: run the checker on your extracted fields. This is a label-free signal that the private study used as an H1 metric.

Reproduce

git clone https://github.com/caiotheodoro/nfse-canvas && cd nfse-canvas
uv sync --extra dev
PYTHONPATH=src uv run python -m nfse_canvas.synth.dataset --out release/nfse-synth --per-family 2500

The same arguments give byte-identical output. The renderer ships its own font (DejaVu Sans, license in src/nfse_canvas/synth/fonts/LICENSE_DEJAVU), so pixels do not depend on the machine.

Limitations

• Synthetic. Clean fonts, generated values, four layouts. Real notes are messier: scans, handwritten marks, dozens of municipal layouts. Nothing here shows that a model trained on this transfers to real notes.
• Four families only. Brazil has thousands of municipal NFS-e layouts. The 2026 national standard (ADN) is not modelled.
• No free-text semantics. Service descriptions come from a short fixed list.
• No baselines are reported for this set yet.

Related work

The pre-registered private study behind this set (same checker, 604 real units, 4 layouts) is in docs/PREREG.md, docs/RESULTS.md and docs/RED-TEAM.md of the repository.

License

Data: CC BY 4.0. Code: see the repository. Font: DejaVu license (bundled).