Thai-Synth-Receipts

Thai-Synth-Receipts is a large-scale, highly robust synthetic dataset of Thai commercial documents designed specifically for training and evaluating state-of-the-art Document AI and Optical Character Recognition (OCR) models. The dataset consists of 14,976 high-resolution document images (Receipts, Tax Invoices, Thermal Slips, and Quotations) across three distinct degradation variants, representing challenging environmental and hardware-induced noise found in real-world document processing.

Dataset Overview

• Language: Thai (with standard English headers)
• Domain: Commercial documents (Supermarket, Cafe, IT Equipment, Construction)
• Samples: 14,976 images (from 4,992 unique documents)
• Task: Optical Character Recognition (OCR), Key Information Extraction (KIE), Document Layout Analysis
• License: MIT
• Source: Procedurally generated via synthetic data engine

Dataset Structure

The dataset is strictly partitioned to prevent data leakage and ensure true zero-shot evaluation capabilities.

Splits

• train: 11,964 images (80%)
• validation: 1,482 images (10%)
• test: 1,530 images (10%)

Data Fields

Each split contains a Hugging Face compatible metadata.jsonl file with the following structure:

• file_name (string): Relative path to the image file
• sample_id (string): Unique identifier for the document
• doc_type (string): Type of document (e.g., tax_invoice, receipt, thermal_slip)
• merchant_name (string): Ground truth store/company name
• items (list): Array of purchased items including name, quantity, and price
• total (float): Total transaction amount
• ocr_boxes (list): Precise polygon coordinates [[x1,y1], [x2,y1], [x2,y2], [x1,y2]] and text transcriptions for every word on the page

Data Statistics

• Total Unique Documents: 4,992
• Variants per Document: 3 (clean, error, photo)
• Total Images: 14,976
• Layout Templates: 6 distinct templates (ranging from structured formal grids to unstructured handwritten lists)

Example

from datasets import load_dataset

dataset = load_dataset("DuckerMaster/Thai-Synth-Receipts")

# Access the train split
train_data = dataset['train']

# View a sample's metadata
print(train_data[0]['merchant_name'])
print(train_data[0]['total'])
print(train_data[0]['ocr_boxes'][0]) # Bounding box and text

Dataset Characteristics

Robustness Evaluation Design

Thai-Synth-Receipts is specifically designed to measure and improve Document AI performance in "in-the-wild" conditions that challenge both optical recognition and layout understanding. Every document is generated in three parallel variants:

1. Clean: Perfect digital baseline with no artifacts.
2. Error (Printer Degradation): Simulates hardware defects including faded ink, vertical streaks, folded paper, ghosting, and ink blobs.
3. Photo (Camera Degradation): Simulates smartphone captures with motion blur, uneven lighting, non-linear gamma curves, and perspective warping.

Zero-Shot Generalization Strategy

Unlike standard random-split datasets, the validation and test sets were curated using an Unseen Dictionary Strategy:

• 100% Disjoint Entities: All merchant names, customer names, addresses, and product items in the validation and test splits are entirely absent from the train split.
• True OCR Evaluation: This forces the model to actually read the Thai text rather than memorizing common vocabulary or exploiting language model hallucinations.

Key Evaluation Aspects

This dataset evaluates Document AI models on:

1. Visual Robustness: Performance on noisy, degraded, and warped document images.
2. Layout Diversity: Ability to parse information across 6 radically different structures (formal grids vs. minimalist POS vs. handwritten formats).
3. Zero-shot Entity Extraction: Accurate transcription of unseen proper nouns and technical terms (e.g., construction materials, IT jargon).

Usage

Loading the Dataset

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("DuckerMaster/Thai-Synth-Receipts")

# Iterate through images and bounding boxes
for sample in dataset['train']:
    image = sample['image'] # PIL Image object
    ocr_boxes = sample['ocr_boxes']
    
    # Your training/evaluation code here

Dataset Creation

Source and Generation

The dataset was generated using a custom procedural synthesis engine in Python. All text entities (names, addresses, tax IDs, phone numbers) were procedurally combined from a large, custom-built Thai vocabulary catalog.

Privacy and Safety

This dataset contains zero Personally Identifiable Information (PII). All companies, individuals, and contact details are 100% fictional. It is completely safe for commercial and academic use without privacy compliance risks.

Limitations

• Synthetic Nature: While visually robust, the dataset is synthetic and may lack certain microscopic textures or artifacts found exclusively in authentic scanned paper.
• Font Diversity: The generation relies on a curated subset of Thai fonts; it may not cover extremely rare or heavily stylized legacy fonts.
• Table Complexity: The dataset focuses on standard commercial transactions and does not include highly complex nested tables or multi-page documents.