Khasi-OCR-81K

This is a large-scale Vision-Language dataset for the Khasi language, specifically curated for OCR, Document Grounding, and Handwriting recognition tasks.

Dataset Summary

• Total Rows: 81,572
• Unique Images: 27,582
• Language: Khasi (kha)

Data Structure

The dataset is provided in sharded Parquet format to ensure stability and compatibility with streaming.

• id: Unique identifier for the task/image pair.
• image: High-quality image data (embedded as bytes).
• conversations: Instruction-tuned format with user and assistant roles.

Tasks Included

1. Markdown Grounding: Structured document conversion with bounding boxes.
2. Free OCR: Plain text extraction from diverse document types.
3. Locate: Identifying and bounding specific text segments.

Usage

Since the dataset is over 70GB, it is highly recommended to use streaming mode to save local disk space.

from datasets import load_dataset

dataset = load_dataset("toiar/Khasi-OCR-81K", split="train", streaming=True)

it = iter(dataset)

# Fetch and display 3 at once
for i in range(3):
    example = next(it)
    
    print(f"\n{'='*60}")
    print(f"EXAMPLE #{i+1} | ID: {example['id']}")
    print(f"{'='*60}")

    for chat in example['conversations']:
        role = chat['role'].upper()
        content = chat['content']
        print(f"[{role}]:")
        print(content)
        print("-" * 30)

    display(example['image'])