invoices-sample / README.md
jkjarvis's picture
Add sample and dataset card
17760ba verified
|
Raw History Blame Contribute Delete
5.95 kB
metadata
license: cc-by-4.0
pretty_name: >-
  Datadhara Sample: Invoice and Receipt Line Items (Handwritten, Printed,
  Thermal)
language:
  - en
task_categories:
  - document-question-answering
  - image-to-text
  - table-question-answering
size_categories:
  - n<1K
tags:
  - document-ai
  - invoices
  - receipts
  - ocr
  - key-value-extraction
  - handwriting
  - line-items
  - de-identified
  - dataset-sample
configs:
  - config_name: default
    data_files:
      - split: train
        path: datadhara-sample-invoices.csv

Invoices and bills: sample

Line-item labels for receipts, bills and invoices: item, quantity, unit, unit price, line total, tax rate, invoice total and document type, with the vendor name redacted. This file shows the schema and labels of the document data that Datadhara licenses to AI teams.

What this is

  • 31 rows: 10 invoices with 2 to 5 line items each, one row per line item.
  • 4 document types: thermal receipt (phone photo), handwritten bill (phone photo), handwritten bill (scan), printed invoice (scan).
  • Buyers are tagged by type only: office, retail shop, clinic, restaurant.
  • Vendor names are [REDACTED]; vendors are identified by a pseudonymous ID.
  • Totals are consistent: for every invoice, invoice_total equals the sum of line_total plus tax at tax_rate_pct, rounded to cents, and every line_total equals qty x unit_price.

The rows in this file were created to show the format, labels and de-identification. They do not come from real businesses, and the consent_ref values are format examples. Licensed deliveries contain real documents collected under written agreements.

Fields

Column Type Description
invoice_no string Invoice or bill number as printed
invoice_date string Invoice date, YYYY-MM-DD
vendor_id string Pseudonymous vendor ID, e.g. vnd_1b993974
vendor_name string Vendor name, always [REDACTED]
buyer_type string office, retail shop, clinic or restaurant
line_no int Line number within the invoice
description string Item description as written, e.g. Rice 25kg bag
qty number Quantity
unit string Unit, e.g. pack, bag, carton, can, case, ream
unit_price number Price per unit
line_total number qty x unit_price
tax_rate_pct number Tax rate for the invoice, in percent (0, 5 or 12 here)
invoice_total number Invoice total including tax (repeated on each line)
currency string Currency code (USD in this sample)
document_type string What the document is and how it was captured
image_file string Name of the matching image, e.g. inv_0001.jpg (images not included in this sample)
consent_ref string Reference to the source agreement and its version, e.g. src-0181 / v2

Example row

{
  "invoice_no": "69803",
  "invoice_date": "2026-09-04",
  "vendor_id": "vnd_91d8ce15",
  "vendor_name": "[REDACTED]",
  "buyer_type": "office",
  "line_no": 2,
  "description": "Wheat flour 10kg",
  "qty": 12,
  "unit": "bag",
  "unit_price": 7.4,
  "line_total": 88.8,
  "tax_rate_pct": 5,
  "invoice_total": 704.24,
  "currency": "USD",
  "document_type": "handwritten bill (phone photo)",
  "image_file": "inv_0003.jpg",
  "consent_ref": "src-0183 / v2"
}

Load it

from datasets import load_dataset

ds = load_dataset("datadhara/invoices-sample", split="train")

Intended uses

Use this sample to check the schema and label set against your pipeline before requesting real data. The licensed data is used for:

  • OCR and key-value extraction
  • Handwriting recognition
  • Line-item table parsing
  • Evaluation sets for document agents

Limitations of this sample

  • Labels only. The image_file column names the scan or photo a delivery would include; no images are in this repository.
  • 10 invoices is enough to see the format, not to train or benchmark anything.
  • Invoice-level fields (invoice_total, tax_rate_pct, document_type) are repeated on each line item.

Full dataset and licensing

This sample file is released under CC BY 4.0, as on the samples page.

The full data is not on Hugging Face. It is available under a commercial licence. Datadhara sources invoices, bills, purchase orders, catalogues and price lists from wholesalers, shops and suppliers, scanned or photographed, printed or handwritten, with line items, quantities, prices, totals and dates in JSON. Bounding boxes can be added if your spec needs them. Volume and labels are agreed per project. Licences cover training and evaluating commercial models, and can be exclusive or non-exclusive.

To request data, describe what you need at datadhara.in/document-data-for-ai or email hello@datadhara.in.

Consent and de-identification

As described in Datadhara's methodology:

  • Every source organisation signs a written agreement that explicitly covers use for AI training. Originals stay with the source.
  • Names, phone numbers, addresses, tax IDs and bank details of vendors and customers are removed from the labels and masked on the images.
  • A sample of each batch is checked against the originals for transcription accuracy, label consistency and anything de-identification missed.
  • Every record carries a consent_ref linking it to its source agreement. In a real delivery, that reference resolves to the signed agreement.

Citation

The same sample files are archived on Zenodo: https://doi.org/10.5281/zenodo.24349905