Khasi-OCR-21K

Khasi-OCR-21K is a curated Vision-Language dataset totaling 21,319 samples, specifically designed to train robust OCR models for the Khasi language. This version introduces a significant amount of high-quality real book data alongside synthetic samples to handle diverse document conditions.

Dataset Split

To ensure reliable model evaluation, the dataset is split into:

• Training Set: ~20,000 samples.
• Validation Set: 1,319 samples (Randomized with a 40/30/30 distribution of New Khasi, Khasi/Synthetic, and English data).

---

Dataset Composition

Source Category	Samples	Description
New Khasi Books	11,319	100 Khasi volumes (moderate scans) Generated via Dak-OCR-v1 and manually refined for maximum accuracy.
Khasi Books + Synthetic	5,000	A mix of 9 Khasi books and synthetic data (including handwritten and obscured conditions).
English Books	5,000	English book scans to maintain multilingual OCR capability.
Total	21,319

---

Image Quality & Types

This dataset provides higher complexity than the previous version by incorporating:

• Scan Quality: clean, moderate (from the new 11K set), and obscured (synthetic).
• Scan Type: Predominantly document with a specific subset of handwriting to improve recognition of non-digital text.

---

Dataset Structure

Each sample follows the DeepSeek-VL/Unsloth formatting for seamless integration into Vision-Language Model training:

• id — Unique identifier (e.g., img_00001).
• image — Embedded pixel data (natively cast).
• scan_quality — clean, moderate, or obscured.
• scan_type — document or handwriting.
• language — kha (Khasi) or en (English).
• messages — Formatted conversation with <|User|> and <|Assistant|> roles.

Conversation Example (Markdown Output)

[
  {
    "role": "<|User|>",
    "content": "<image>\nFree OCR.",
    "images": ["images/img_00001.jpg"]
  },
  {
    "role": "<|Assistant|>",
    "content": "# KA JINGLAMPHRANG\n\nExtracted text with preserved markdown structure..."
  }
]

---

Technical Tasks

• Free OCR: Direct transcription of document images into structured Markdown.
• Layout Preservation: Training the model to recognize and replicate headings, lists, and tabular data from the original image.
• Handwriting Recognition: Improved accuracy on cursive and handwritten Khasi notes through the synthetic handwriting subset.