sinabis-group/sinabis-1m-synthetic-pdf

Sinabis Green Energy – Synthetic Document Dataset (1M PDFs)

More than 1,006,000 business documents (PDF) from the everyday operations of a fictional renewable-energy group (Sinabis Green Energy) and its worldwide ecosystem of customers, suppliers, subcontractors, utilities, authorities, insurers, banks, hotels and service providers.

Every person, company, address, phone number, bank account, tax number, invoice and transaction in this dataset is purely fictional. The underlying world and business taxonomy were created with LLM assistance, while all documents, calculations, serial numbers, dates and visual company designs were rendered deterministically using ReportLab and custom vector engines.

All figures in this dataset card were computed directly from the generated files (as of 16 Oct 2026). They are not target estimates.

At a glance

Metric	Value
Total Documents	1,006,000 PDFs (71.8 GB uncompressed, 65.9 GB in 7 ZIP archives)
Document Types	20 distinct business themes (invoices, receipts, payslips, bank statements, certificates ...)
Vector vs. Scanned	645,457 clean vector PDFs (64.2 %) · 360,543 scanned/photocopied (35.8 %) with stamps, handwritten notes, skew and sensor noise
Time span	2014-03-14 to 2026-06-30 (12.3 years of continuous commercial history)
Languages	24 languages (EN 83.1 %, ZH 4.2 %, rest across 22 languages)
Issuing Companies	1,777 distinct fictional companies across 33 countries
Unique Visual Identities	178,099 individual company Design-DNA signatures

Quick start

The Parquet files under data/ hold one row per document: all manifest fields plus the extracted text, so the dataset can be browsed in the Dataset Viewer above and loaded directly:

from datasets import load_dataset

ds = load_dataset("sinabis-group/sinabis-1m-synthetic-pdf", split="train")
print(ds[0]["theme"], ds[0]["number"], ds[0]["total"], ds[0]["currency"])
print(ds[0]["text"])

vector = ds.filter(lambda r: not r["scanned"])   # documents with a text layer

The text column holds the text layer of the PDF, page by page; pages is the page count. The scanned/photocopied documents (scanned == True) are images without a text layer, so their text is empty: they are meant for OCR and vision models. The original PDFs are in the ZIP archives.

Downloads

The dataset is packaged in rolling ~10 GB ZIP archives for fast parallel downloads and streaming:

File	Size	Contents
data/train-*.parquet (11 files)	458 MB	Manifest fields + extracted text, one row per document (Dataset Viewer, load_dataset)
Sinabis_1M_pdfs_part_01.zip	9.86 GB	150,500 PDFs (10.74 GB uncompressed)
Sinabis_1M_pdfs_part_02.zip	9.88 GB	151,000 PDFs (10.76 GB uncompressed)
Sinabis_1M_pdfs_part_03.zip	9.89 GB	151,000 PDFs (10.77 GB uncompressed)
Sinabis_1M_pdfs_part_04.zip	9.87 GB	151,500 PDFs (10.74 GB uncompressed)
Sinabis_1M_pdfs_part_05.zip	9.89 GB	150,500 PDFs (10.77 GB uncompressed)
Sinabis_1M_pdfs_part_06.zip	9.88 GB	150,500 PDFs (10.76 GB uncompressed)
Sinabis_1M_pdfs_part_07.zip	6.65 GB	101,000 PDFs (7.24 GB uncompressed)
Sinabis_1M_pdfs_manifest.zip	74.3 MB	Complete dataset index: 1,006,000 JSONL records with file paths, metadata and sums

Unpacking

All archives share a unified directory structure. Unpacking them into the same directory merges all files seamlessly:

# Download and unpack with the Hugging Face CLI
hf download sinabis-group/sinabis-1m-synthetic-pdf --repo-type dataset --local-dir sinabis-1m-pdf
cd sinabis-1m-pdf
for z in Sinabis_1M_pdfs_*.zip; do unzip -q "$z"; done

Folder structure

In the repository:

.
├── README.md
├── data/
│   └── train-*.parquet                 (rows for the Dataset Viewer and load_dataset)
├── Sinabis_1M_pdfs_part_01.zip … part_07.zip    → pdf/ (+ manifest/ shards)
└── Sinabis_1M_pdfs_manifest.zip                 → manifest/

After unpacking:

sinabis-1m-pdf/
├── README.md
├── pdf/
│   ├── invoice/
│   │   ├── 0000/
│   │   │   ├── 0000000_INV-40192.pdf
│   │   │   └── ...
│   │   └── ...
│   ├── payslip/
│   ├── utility_bill/
│   └── ... (20 themes)
└── manifest/
    ├── b00000.jsonl
    ├── b00001.jsonl
    └── ...

Each bNNNNN.jsonl manifest contains exact metadata for every document:

• file: Relative path to the PDF
• theme: Document theme (e.g. invoice, payslip)
• issuer: Fictional issuing company name
• issuer_country: Two-letter country code (ISO 3166-1)
• recipient: Recipient name
• lang: Language code (ISO 639-1)
• date: Document issue date (ISO 8601)
• number: Serial document number
• currency: ISO 4217 currency code
• total: Document net/gross total (recalculable from line items)
• scanned: Boolean flag indicating whether physical scanner simulation was applied
• bytes: Exact byte size

Document Themes

Theme	Documents	Share	Description
invoice	281,680	28.0 %	Trade and service invoices with tax and early-payment discount calculations
receipt	90,540	9.0 %	Receipts and till slips for cash and card payments
quote	60,360	6.0 %	Commercial quotations with validity periods and payment terms
payslip	50,300	5.0 %	Monthly payslips with deductions, taxes and social security contributions
delivery_note	50,300	5.0 %	Delivery notes with package details and handover receipt
bank_statement	40,240	4.0 %	Bank statements with transaction descriptions and running balances
utility_bill	40,240	4.0 %	Energy and utility bills with consumption history
certificate	40,240	4.0 %	Participation, training and calibration certificates with seal
dunning	40,240	4.0 %	Multi-level payment reminders with dunning fees and deadlines
order_confirmation	40,240	4.0 %	Order confirmations with delivery dates
timesheet	30,180	3.0 %	Monthly timesheets with project and daywork hours
freight_bill	30,180	3.0 %	Freight and forwarding documents (CMR/AWB) with dangerous-goods notes
hotel_invoice	30,180	3.0 %	Hotel invoices with tourist tax, lodging and catering
authority_notice	30,180	3.0 %	Authority notices and fee decisions with appeal instructions
maintenance_report	30,180	3.0 %	Technical maintenance and inspection reports with test points
workshop_invoice	30,180	3.0 %	Workshop invoices with labour units (AW) and spare parts
expense_report	30,180	3.0 %	Travel expense reports with meal allowances and receipts
insurance_policy	20,120	2.0 %	Insurance policies with coverage amounts and clauses
packing_list	20,120	2.0 %	Detailed packing lists with weight and dimension details
credit_note	20,120	2.0 %	Credit notes referencing the original invoice

Language Distribution

Language	Documents	Share
English (en)	836,257	83.1 %
Chinese (zh)	42,399	4.2 %
Japanese (ja)	40,678	4.0 %
German (de)	16,166	1.6 %
Spanish (es)	8,659	0.9 %
Italian (it)	6,710	0.7 %
Portuguese (pt)	6,616	0.7 %
French (fr)	4,945	0.5 %
Swedish (sv)	4,882	0.5 %
Polish (pl)	4,808	0.5 %
Dutch (nl)	4,069	0.4 %
Korean (ko)	3,684	0.4 %
Hungarian (hu)	3,417	0.3 %
Norwegian (nb)	3,043	0.3 %
Danish (da)	2,713	0.3 %
Vietnamese (vi)	2,575	0.3 %
Greek (el)	2,351	0.2 %
Indonesian (id)	2,278	0.2 %
Turkish (tr)	2,005	0.2 %
Finnish (fi)	1,840	0.2 %
Czech (cs)	1,733	0.2 %
Ukrainian (uk)	1,609	0.2 %
Romanian (ro)	1,541	0.2 %
Russian (ru)	1,022	0.1 %

Visual Design-DNA

Every issuing company in the dataset possesses a deterministic visual identity derived from its unique company ID:

• Header Archetypes: Full-bleed color band, clean letterhead, modern sidebar, minimalist, split-block.
• Procedural Logos: 20 vector geometric symbols, typography and layouts generated with exact brand rules.
• Color Palettes: Harmonized primary, secondary, accent and neutral tones with WCAG-compliant contrast.
• Typography: Tailored font pairings for the scripts used by the 24 languages (Latin, CJK, Cyrillic and Greek).
• Physical Scan Artifacts: About a third (35.8 %) of the documents undergo a high-resolution copier/scanner simulation featuring ink stamps, ballpoint pen annotations, checkmarks, paper fiber textures, tilt, and optical noise.

Strict Fictionality Guardrails

• Domain Names: All external email and website domains use the RFC 2606 reserved .example TLD.
• Phone Numbers: All telephone numbers use fictional prefixes (e.g. 555-01xx in North America).
• Banking: Bank names and IBANs are algorithmically generated using synthetic syllables and valid checksum algorithms.
• Screening: Every entity is filtered against a 5,000+ entry real-world trademark and organization blocklist (guard.json).

License

This dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license. You are free to share and adapt the material for any purpose, including commercial applications, provided appropriate credit is given:

Sinabis GmbH (2026). Sinabis Green Energy – Synthetic Document Dataset (1M PDFs). https://huggingface.co/datasets/sinabis-group/sinabis-1m-synthetic-pdf

Citation

@misc{sinabis_green_pdf_dataset_2026,
  author = {{Sinabis GmbH}},
  title  = {Sinabis Green Energy -- Synthetic Document Dataset (1M PDFs)},
  year   = {2026},
  url    = {https://huggingface.co/datasets/sinabis-group/sinabis-1m-synthetic-pdf}
}