README.md

metadata

dataset_info:
  features:
    - name: image
      dtype: image
    - name: text
      dtype: string
  splits:
    - name: train
      num_bytes: 508398188
      num_examples: 441
  download_size: 506925962
  dataset_size: 508398188
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
license: cc-by-4.0
task_categories:
  - image-to-text
language:
  - sa
tags:
  - ocr
size_categories:
  - n<1K

Sanskrit Dataset

Overview

This dataset is a conversion of the Pracalit for Sanskrit and Newar MSS 16th to 19th C., Ground Truth dataset, originally published on Zenodo. The dataset contains pairs of images and corresponding plain text extracted from XML files. This dataset specifically includes the ground truth data from the original repository.

Dataset Description

• Images: The original images from the dataset.
• Text: The text is extracted from XML files using the PAGE XML schema.

Potential Use for VLM-based OCR Training

The existing ground truth OCR data can be particularly useful for bootstrapping datasets for Vision Language Model (VLM) based OCR training. By leveraging the high-quality annotations provided in this dataset, researchers can train models to recognize and transcribe text from images more accurately. This can be especially beneficial for languages and scripts that are underrepresented in existing OCR datasets.

Processing Script

The dataset was processed using the following script:

import os
import xml.etree.ElementTree as ET
from pathlib import Path
from PIL import Image as PILImage
import datasets
from datasets import Image, Features, Dataset
from tqdm import tqdm

def extract_text_from_xml(xml_path):
    """Extract plain text from XML file."""
    tree = ET.parse(xml_path)
    root = tree.getroot()
    text_lines = []
    for textline in root.findall('.//{http://schema.primaresearch.org/PAGE/gts/pagecontent/2013-07-15}TextLine'):
        text = textline.find('.//{http://schema.primaresearch.org/PAGE/gts/pagecontent/2013-07-15}TextEquiv/{http://schema.primaresearch.org/PAGE/gts/pagecontent/2013-07-15}Unicode')
        if text is not None and text.text:
            text_lines.append(text.text.strip())
    return ' '.join(text_lines)

def create_dataset():
    base_dir = Path("Copy_of_HTR_Train_Set_'Pracalit_for_Sanskrit_and_Newar_MSS_16th_to_19th_C_'")
    page_dir = base_dir / "page"
    images = []
    texts = []
    image_files = list(base_dir.glob("*.jpg"))
    for img_file in tqdm(image_files, desc="Processing images"):
        xml_file = page_dir / f"{img_file.stem}.xml"
        if not xml_file.exists():
            print(f"Warning: No matching XML found for {img_file.name}")
            continue
        try:
            text = extract_text_from_xml(xml_file)
            img = PILImage.open(img_file)
            img.verify()
            images.append(str(img_file))
            texts.append(text)
        except Exception as e:
            print(f"Error processing {img_file.name}: {str(e)}")
            continue
    dataset = Dataset.from_dict({
        "image": images,
        "text": texts
    }, features=Features({
        "image": Image(),
        "text": datasets.Value("string")
    }))
    print(f"Dataset created with {len(dataset)} examples")
    print("Saved to sanskrit_dataset.parquet")
    dataset.push_to_hub("davanstrien/sanskrit_dataset", private=True)

if __name__ == "__main__":
    create_dataset()

Citation

If you use this dataset, please cite the original dataset on Zenodo: OCR model for Pracalit for Sanskrit and Newar MSS 16th to 19th C., Ground Truth