dh-unibe/image-text_kurrent-xix

Dataset Card for transkribus-exports-30840-raw-xml

This dataset was created using pagexml-hf converter from Transkribus PageXML data.

Dataset Summary

This dataset contains 137487 samples across 1 split(s). All of the images and the data concern images and transcriptions in German from the 19th century.

The dataset was used to create the linked model TrOCR: (https://huggingface.co/dh-unibe/trocr-kurrent).

For an evaluation see: Hodel, Tobias, David Schoch, Christa Schneider, and Jake Purcell. “General Models for Handwritten Text Recognition: Feasibility and State-of-the Art. German Kurrent as an Example.” Journal of Open Humanities Data 7 (July 2021): 13. (https://doi.org/10.5334/johd.46).

Dataset Includes

Text from:

• University Archives of Greifswald: Konzilsprotokolle (see also: https://ariadne-portal.uni-greifswald.de/?arc=1&type=obj&id=5164866): -- A_Haeckermann
• GTA ETHZ: -- Parts of the GT produced within the Scholarly edition "Semper" (https://www.semper-edition.ch/)

Dataset Structure

Data Splits

• train: 137487 samples

Dataset Size

• Approximate total size: 994516.35 MB
• Total samples: 137487

Features

• image: Image(mode=None, decode=False)
• xml_content: Value('string')
• filename: Value('string')
• project_name: Value('string')

Data Organization

Data is organized as parquet shards by split and project:

data/
├── <split>/
│   └── <project_name>/
│       └── <timestamp>-<shard>.parquet

The HuggingFace Hub automatically merges all parquet files when loading the dataset.

Usage

from datasets import load_dataset

# Load entire dataset
dataset = load_dataset("dh-unibe/transkribus-exports-30840-raw-xml") 

# Load specific split
train_dataset = load_dataset("dh-unibe/transkribus-exports-30840-raw-xml", split="train")