The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.
YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
Receipts, Invoices & Bank Statements Dataset
Dataset Overview
The Receipts, Invoices & Bank Statements Dataset is a large-scale document understanding dataset containing 150,000 document images paired with structured JSON annotations.
The dataset contains three major document categories:
- 50,000 Receipts
- 50,000 Invoices
- 50,000 Bank Statements
The dataset is designed for research and development in:
- Document AI
- OCR
- Document Understanding
- Information Extraction
- Structured Data Extraction
- Multimodal AI
- Vision-Language Models (VLMs)
- Image-to-JSON generation
- Financial document processing
- Document classification
- Document information extraction
Dataset Statistics
| Document Type | Number of Samples |
|---|---|
| Receipts | 50,000 |
| Invoices | 50,000 |
| Bank Statements | 50,000 |
| Total | 150,000 |
Dataset Size
Approximate compressed dataset size: 33.2 GB
Dataset Composition
Total Dataset
|
|-- Receipts
| `-- 50,000 documents
|
|-- Invoices
| `-- 50,000 documents
|
`-- Bank Statements
`-- 50,000 documents
Dataset Details
Dataset Description
This dataset provides a large collection of document images and corresponding structured JSON annotations.
Each sample consists conceptually of:
Document Image
+
Structured JSON Annotation
The dataset is intended to help researchers and developers build systems capable of understanding semi-structured financial and commercial documents.
The primary objective is to support systems that transform document images into machine-readable structured information.
A typical processing pipeline is:
Document Image
|
v
Image Preprocessing
|
v
OCR / Vision Encoder
|
v
Document Understanding
|
v
Information Extraction
|
v
Structured JSON
Document Categories
Receipts
The dataset contains 50,000 receipt images with corresponding JSON annotations.
Receipt information may include:
- Merchant name
- Merchant address
- Receipt number
- Invoice number
- Date
- Time
- Currency
- Items
- Product names
- Quantity
- Unit price
- Item total
- Discount
- Tax
- Subtotal
- Grand total
- Payment information
The exact fields depend on the annotation schema used for each sample.
Invoices
The dataset contains 50,000 invoice images with corresponding JSON annotations.
Invoice information may include:
- Invoice number
- Invoice date
- Due date
- Seller information
- Buyer information
- Billing information
- Shipping information
- Product or service details
- Quantity
- Unit price
- Item total
- Discount
- Tax
- Subtotal
- Grand total
- Currency
- Payment information
The exact fields depend on the annotation schema used for each sample.
Bank Statements
The dataset contains 50,000 bank statement images with corresponding JSON annotations.
Bank statement information may include:
- Account holder information
- Account information
- Statement period
- Opening balance
- Closing balance
- Transaction date
- Transaction description
- Debit
- Credit
- Transaction amount
- Running balance
- Transaction reference
The exact fields depend on the annotation schema used for each sample.
Dataset Structure
The recommended repository structure is:
receipts-invoices-bank-statements/
|
|-- README.md
|-- LICENSE
|-- dataset_summary.json
|
|-- receipts/
| |
| |-- train/
| | |-- images/
| | `-- metadata.jsonl
| |
| |-- validation/
| | |-- images/
| | `-- metadata.jsonl
| |
| `-- test/
| |-- images/
| `-- metadata.jsonl
|
|-- invoices/
| |
| |-- train/
| | |-- images/
| | `-- metadata.jsonl
| |
| |-- validation/
| | |-- images/
| | `-- metadata.jsonl
| |
| `-- test/
| |-- images/
| `-- metadata.jsonl
|
`-- bank_statements/
|
|-- train/
| |-- images/
| `-- metadata.jsonl
|
|-- validation/
| |-- images/
| `-- metadata.jsonl
|
`-- test/
|-- images/
`-- metadata.jsonl
Dataset Splits
The recommended split is an 80/10/10 train-validation-test split.
| Document Type | Train | Validation | Test | Total |
|---|---|---|---|---|
| Receipts | 40,000 | 5,000 | 5,000 | 50,000 |
| Invoices | 40,000 | 5,000 | 5,000 | 50,000 |
| Bank Statements | 40,000 | 5,000 | 5,000 | 50,000 |
| Total | 120,000 | 15,000 | 15,000 | 150,000 |
Important: The split numbers above should only be considered final if these splits have actually been created in the repository.
For reliable evaluation, documents originating from the same source should remain in the same split to reduce the possibility of data leakage.
Annotation Format
Each document image is associated with structured JSON information.
The general relationship is:
Image
+
JSON Label
Example Receipt Annotation
{
"document_type": "receipt",
"merchant": {
"name": "Example Store",
"address": "Example Address"
},
"receipt_number": "REC-001",
"date": "2026-01-15",
"currency": "INR",
"items": [
{
"name": "Example Product",
"quantity": 2,
"unit_price": 100,
"total": 200
}
],
"subtotal": 200,
"tax": 36,
"discount": 0,
"grand_total": 236
}
Example Invoice Annotation
{
"document_type": "invoice",
"invoice_number": "INV-001",
"invoice_date": "2026-01-15",
"due_date": "2026-02-15",
"seller": {},
"buyer": {},
"items": [],
"subtotal": 1000,
"tax": 180,
"discount": 0,
"grand_total": 1180
}
Example Bank Statement Annotation
{
"document_type": "bank_statement",
"statement_period": {
"start": "2026-01-01",
"end": "2026-01-31"
},
"transactions": [
{
"date": "2026-01-05",
"description": "Example Transaction",
"debit": 500,
"credit": 0,
"balance": 9500
}
]
}
The examples above are illustrative. The actual JSON schema in the dataset should be treated as the authoritative annotation format.
Metadata Format
For Hugging Face-compatible image datasets, metadata can be represented using metadata.jsonl.
Example:
{"file_name":"images/receipt_000001.jpg","document_type":"receipt","label":{"merchant":"Example Store","date":"2026-01-15","total":236}}
{"file_name":"images/receipt_000002.jpg","document_type":"receipt","label":{"merchant":"Example Market","date":"2026-01-16","total":850}}
The file_name field associates each metadata record with its corresponding image.
Intended Uses
This dataset is intended for:
- Research
- Development
- Experimentation
- Benchmarking
- Education
- Document AI development
- OCR research
- Multimodal AI research
Document AI
Potential applications include:
- Receipt understanding
- Invoice understanding
- Bank statement understanding
- Document classification
- Document type detection
- Document information extraction
- Financial document processing
- Commercial document processing
OCR
Potential applications include:
- OCR evaluation
- Text recognition
- Text extraction
- OCR post-processing
- OCR error correction
- Reading-order analysis
- Document text recognition
Information Extraction
Potential tasks include:
- Key-value extraction
- Entity extraction
- Financial information extraction
- Invoice field extraction
- Receipt field extraction
- Transaction extraction
- Line-item extraction
- Table extraction
- Structured JSON generation
Multimodal AI
Potential applications include:
- Image-to-JSON generation
- Document VLM training
- Document VLM evaluation
- Vision-language model benchmarking
- Multimodal instruction tuning
- Document question answering
- Document visual reasoning
Potential Research Tasks
Document Classification
Document Image
|
v
Receipt / Invoice / Bank Statement
OCR
Document Image
|
v
Machine-readable Text
Information Extraction
Document Image
|
v
Entities and Fields
Structured Data Extraction
Document Image
|
v
Structured JSON
Multimodal Learning
Image + Text Prompt
|
v
Structured Response
Document VLM Evaluation
Document Image
|
v
Vision-Language Model
|
v
Structured JSON
Example AI Pipeline
A typical application using this dataset may follow:
DOCUMENT IMAGE
|
v
Image Preprocessing
|
v
OCR / VLM
|
v
Document Representation
|
v
Information Extraction
|
v
Structured JSON
For example:
Receipt Image
|
v
OCR
|
v
Extracted Text
|
v
LLM / VLM
|
v
Receipt JSON
Out-of-Scope Uses
This dataset should not be used for:
- Fraudulent financial activity
- Identity theft
- Unauthorized financial profiling
- Unauthorized processing of personal financial information
- Creating fraudulent financial documents
- Creating fake documents for illegal activities
- Deceiving financial institutions
- Unauthorized inference of sensitive personal information
- Making high-impact financial decisions about individuals without appropriate validation
The dataset should not be considered authoritative financial information.
Models trained using this dataset should be independently validated before deployment in production or high-impact financial applications.
Dataset Creation
Curation Rationale
The dataset was created to provide a large-scale collection of document images and structured annotations for developing and evaluating modern Document AI and multimodal information extraction systems.
The dataset combines three important document categories:
- Receipts
- Invoices
- Bank Statements
The unified structure enables researchers and developers to experiment with:
- Document classification
- OCR
- Information extraction
- Document understanding
- Image-to-JSON generation
- Multimodal model evaluation
- Financial document processing
Source Data
The exact source-data methodology should be documented according to the final dataset creation process.
The dataset may contain documents originating from:
- Synthetic document generation
- Publicly available datasets
- Manually created documents
- Other authorized document sources
If documents originate from third-party datasets or publicly available sources, the original datasets and their licensing conditions should be documented.
If documents originate from real-world financial or commercial documents, appropriate authorization and redistribution rights should be verified.
Data Collection and Processing
The general processing pipeline is:
Source / Generated Document
|
v
Document Image
|
v
Image Quality Checking
|
v
Document Categorization
|
v
JSON Annotation
|
v
Image + JSON Pairing
|
v
Validation
|
v
Dataset Organization
|
v
Hugging Face Repository
Recommended quality-control checks include:
- Image readability
- Image corruption detection
- JSON syntax validation
- Image/JSON pairing validation
- Document type validation
- Required-field validation
- Duplicate detection
- Invalid-value detection
- Annotation consistency checks
- Train/test leakage checks
- Sensitive-information checks
Source Data Producers
The source-data producers depend on how individual documents were created or collected.
If documents are synthetically generated, the generation methodology and software used should be documented.
If documents originate from third-party datasets, the original datasets and their licenses should be documented.
If documents originate from real-world financial or commercial documents, appropriate authorization and redistribution rights should be verified.
Annotation Process
The dataset contains structured JSON annotations corresponding to document images.
The final release should document whether annotations were:
- Manually created
- Programmatically generated
- Automatically extracted
- Generated using OCR
- Generated using AI or LLM systems
- Human verified
- Automatically validated
- Corrected manually
If automated or synthetic annotation was used, this should be explicitly disclosed.
Annotators
If human annotators were involved, the dataset documentation should describe:
- Number of annotators
- Annotation responsibilities
- Annotation guidelines
- Quality-control process
- Validation process
- Inter-annotator agreement, if available
If annotations were generated programmatically, they should be described as machine-generated or programmatically generated annotations.
Personal and Sensitive Information
Financial documents can potentially contain sensitive information.
Examples include:
- Names
- Addresses
- Telephone numbers
- Email addresses
- Bank account numbers
- Payment information
- Transaction information
- Invoice information
- Customer information
- Business information
- Financial amounts
Before public distribution, the dataset should be checked for personally identifiable information (PII), financial information, and other sensitive information.
If real-world documents are included, appropriate authorization, anonymization, redaction, and redistribution rights should be established.
If the dataset is entirely synthetic, this section should explicitly state that the documents are synthetic and do not correspond to real individuals, accounts, businesses, or transactions.
Users should not attempt to recover, infer, or reconstruct redacted or anonymized information.
Bias, Risks, and Limitations
Document Diversity
The visual appearance of the documents may not represent the full diversity of real-world receipts, invoices, and bank statements.
Potential variations include:
- Document layout
- Font
- Resolution
- Image quality
- Lighting
- Background
- Tables
- Language
- Currency
- Document length
- Printing quality
- Scanning quality
- Camera distortion
Geographic Bias
If the dataset is concentrated in a particular country or region, models trained on it may not generalize well to documents from other regions.
Language Bias
If some languages occur more frequently than others, model performance may vary across languages.
The exact language distribution should be measured and documented.
Layout Bias
Models may learn characteristics of frequently occurring layouts rather than general document understanding.
Performance may therefore decrease on previously unseen layouts.
Annotation Errors
Annotations may contain:
- Missing fields
- Incorrect values
- Formatting inconsistencies
- OCR errors
- Incorrect table structures
- Incorrect reading order
Users should validate annotations before using them for benchmarking or production training.
Synthetic Data Limitations
If synthetic documents are included, they may not fully reproduce real-world:
- Printing artifacts
- Camera distortion
- Scanning noise
- Handwriting
- Lighting conditions
- Paper damage
- Compression artifacts
- Unusual document layouts
Recommendations
Users should:
- Validate the dataset before training production models.
- Check image and JSON consistency.
- Remove duplicate documents.
- Validate JSON annotations.
- Use document-level train/validation/test splitting.
- Evaluate each document category separately.
- Evaluate models on external datasets.
- Test models on unseen document layouts.
- Measure performance across languages where applicable.
- Perform human review for high-risk financial applications.
- Properly protect sensitive information.
- Avoid using model predictions as the sole basis for financial decisions.
Recommended Evaluation Metrics
OCR
- Character Error Rate (CER)
- Word Error Rate (WER)
- Text accuracy
Document Classification
- Accuracy
- Precision
- Recall
- F1 score
- Confusion matrix
Information Extraction
- Entity precision
- Entity recall
- Entity F1
- Field-level accuracy
- Key-value extraction accuracy
Structured JSON Extraction
- JSON validity rate
- Exact-match accuracy
- Field-level accuracy
- Nested-field accuracy
- Numerical-value accuracy
- Precision
- Recall
- F1 score
Table Extraction
- Cell-level accuracy
- Row accuracy
- Column accuracy
- Table structure accuracy
Dataset Quality Checklist
Before a production or benchmark release, the following checks are recommended:
[✓] Image exists
[✓] Image can be opened
[✓] Image is not corrupted
[✓] JSON exists
[✓] JSON is valid
[✓] Image and JSON correspond correctly
[✓] Document category is correct
[✓] Required fields are present
[✓] Numerical values are valid
[✓] Duplicate documents are checked
[✓] Train/test leakage is checked
[✓] Sensitive information is checked
[✓] Annotation consistency is checked
[✓] Dataset statistics are verified
Dataset Version
Version 1.0.0
Initial dataset release containing:
- 150,000 total document samples
- 50,000 receipts
- 50,000 invoices
- 50,000 bank statements
- Image and JSON annotation pairs
Future versions should document:
- Added documents
- Removed documents
- Corrected annotations
- Schema changes
- Processing changes
- Quality improvements
- New document categories
- New languages
- New metadata
Future Improvements
Future versions may include:
- Additional document categories
- Additional languages
- Additional document layouts
- Improved JSON annotations
- OCR transcriptions
- Word-level bounding boxes
- Entity bounding boxes
- Table annotations
- Layout annotations
- Question-answer pairs
- Document-level metadata
- Instruction-tuning formats
- Benchmark-specific evaluation splits
- Human-verified annotations
Dataset Loading
Once the repository is organized into a Hugging Face-compatible dataset structure, users may load the dataset using the Hugging Face datasets library.
from datasets import load_dataset
dataset = load_dataset(
"devpatel18042004/receipts-invoices-bank-statements"
)
print(dataset)
The exact loading behavior depends on the final repository structure and supported file formats.
Example Usage
A document-processing application can conceptually use the dataset as:
Image
|
v
Vision Encoder / OCR
|
v
Document Representation
|
v
Information Extraction Model
|
v
JSON Output
License
License: To be specified by the dataset maintainer.
The final license should be selected only after verifying:
- Dataset ownership
- Source-data licenses
- Redistribution rights
- Third-party dataset restrictions
- Synthetic-data licensing
- Commercial-use permissions
If the dataset contains material derived from other datasets, the licenses of those source datasets must also be respected.
Citation
If you use this dataset in research, publications, demonstrations, benchmarks, or applications, please cite the dataset repository.
BibTeX
@dataset{patel_receipts_invoices_bank_statements,
author = {Patel, Dev},
title = {Receipts, Invoices and Bank Statements Dataset},
year = {2026},
publisher = {Hugging Face},
version = {1.0.0}
}
APA
Patel, D. (2026). Receipts, Invoices and Bank Statements Dataset (Version 1.0.0). Hugging Face.
Glossary
Receipt
A document that records a completed purchase or transaction.
Invoice
A commercial document that records products or services supplied and the amount payable.
Bank Statement
A document containing account information and financial transactions over a specified period.
OCR
Optical Character Recognition technology used to convert text contained in images into machine-readable text.
Document AI
Artificial intelligence techniques designed to process, understand, and extract information from documents.
Document Understanding
The process of interpreting the structure, content, and semantic meaning of a document.
Information Extraction
The process of identifying and extracting structured information from unstructured or semi-structured documents.
VLM
Vision-Language Model capable of processing visual and textual information.
JSON
JavaScript Object Notation, a machine-readable structured data format used for representing annotations in this dataset.
PII
Personally Identifiable Information that can potentially identify an individual.
More Information
This dataset is intended to support development and evaluation of modern document-processing systems.
Potential future versions may include additional document types, languages, annotations, OCR data, bounding boxes, tables, question-answer pairs, and multimodal instruction-tuning formats.
Dataset Card Authors
Dev Patel
Maintainer
Dev Patel
Dataset repository:
devpatel18042004/receipts-invoices-bank-statements
Contact
For questions, corrections, dataset issues, or discussions, please use the discussion and issue features available on the Hugging Face dataset repository.
Disclaimer
This dataset is provided for research, development, experimentation, benchmarking, and educational purposes.
Users are responsible for verifying the accuracy, legality, licensing, privacy, and suitability of the dataset for their intended use.
Particular care should be taken when using this dataset for financial-document processing or applications involving sensitive information.
The dataset should not be considered a source of authoritative financial information.
Models trained using this dataset should not be deployed in high-impact financial applications without appropriate validation, monitoring, security controls, and human oversight.
- Downloads last month
- 75