OCR Invoice
Description
This dataset contains annotated receipt transaction data for developing and evaluating information extraction models from Optical Character Recognition (OCR) outputs. The dataset is designed for structured receipt parsing, where noisy OCR text is transformed into standardized JSON representations containing purchased items and transaction charges. The dataset consists of three subsets: 800 training samples, 90 testing samples, and analysis files for performance evaluation. Each training sample includes an OCR-derived word sequence as the input and a manually verified JSON annotation as the ground truth. The JSON schema contains two primary objects: items, which records product names, quantities, and prices, and charge, which includes tax, service charge, discount, and other transaction-level charges. The receipts were collected from various restaurants, cafés, bakeries, retail stores, and food outlets in Indonesia. Consequently, the dataset represents diverse receipt layouts, typography, and transaction formats. The OCR outputs also contain realistic recognition errors such as misspelled words, merged tokens, missing characters, inconsistent number formatting, and layout-related noise, providing a challenging benchmark for document understanding and information extraction research. The testing dataset contains unseen receipt samples together with model predictions and evaluation metrics, enabling reproducible benchmarking of receipt information extraction methods. Additional analysis files provide detailed experimental results, similarity scores, execution time, computational cost, confusion matrices, and performance summaries. This dataset can support research in document intelligence, receipt understanding, OCR post-processing, information extraction, structured document parsing, large language models (LLMs), multimodal artificial intelligence, and intelligent financial document processing.