Synthetic Packaging Slip Donut Dataset

Synthetic packaging slip dataset for Donut fine-tuning.

Contents

• train: 50 PNG images + labels
• val: 50 PNG images + labels
• test: 50 PNG images + labels
• Total: 150 images

Each split has:

• metadata.jsonl: Donut/imagefolder-ready metadata. ground_truth is a JSON string containing gt_parse.
• annotations_with_sequences.jsonl: structured JSON plus a pre-built Donut target sequence.
• annotations.json: pretty JSON version for inspection.

Schema

{
  "packaging": {
    "seller": {"company_name": "", "address": "", "phone_number": ""},
    "buyer": {"bill_to_name": "", "bill_to_address": "", "ship_to_name": "", "ship_to_address": ""},
    "document": {"package_number": "", "order_date": "", "sales_order_number": "", "po_number": ""},
    "items": [{"sr_no": "", "item_description": "", "sku": "", "quantity": "", "unit_price": "", "total_price": ""}],
    "summary": {"total_quantity": "", "sub_total": "", "tax": "", "freight": "", "grand_total": ""},
    "notes": ""
  }
}

Load with Hugging Face datasets

from datasets import load_dataset

dataset = load_dataset("imagefolder", data_dir="packaging_slip_donut_dataset")
print(dataset)

All records are synthetic. Totals are validated during generation.