nn-auto-bench-ds

nn-auto-bench-ds is a dataset designed for key information extraction (KIE) and serves as a benchmark dataset for nn-auto-bench.

Dataset Overview

The dataset comprises 1,000 documents, categorized into the following types:

1. Invoice
2. Receipt
3. Passport
4. Bank Statement

The documents are primarily available in English, with some also in German and Arabic. Each document is annotated for key information extraction and specific tasks. The dataset can be used to compute LLM's oneshot performance on KIE tasks.

Dataset Schema

The dataset includes the following columns:

• image_path: File path to the document image.
• content: OCR-extracted text from the image.
• accepted: Ground truth answer.
• Queried_labels: Labels, fields, or keys targeted for extraction.
• Queried_col_headers: Column headers targeted for extraction.
• ctx_1: OCR text from an example document.
• ctx_1_image_path: File path to the example document’s image.
• ctx_1_accepted: Ground truth answer for the example document.

There are total 54 unique fields/keys/labels that we want to extract from the documents.

Loading the Dataset

To load the dataset in Python using the datasets library:

from datasets import load_dataset

dataset = load_dataset("nanonets/nn-auto-bench-ds")

Data Sources

This dataset aggregates information from multiple open-source datasets, including:

1. German Invoices Dataset
2. Personal Financial Dataset for India
3. RVL-CDIP Invoice Dataset
4. FATURA Dataset
5. Find It Again
6. Generated USA Passports Dataset
7. Synthetic Passports Dataset

This dataset is valuable for benchmarking key information extraction models and advancing research in document understanding and natural language processing (NLP).