Khasi-OCR-36K

Khasi-OCR-36K is a Vision-Language dataset designed for OCR, document understanding, and handwriting recognition in the Khasi language, with a smaller subset of English samples.

This specific version of the dataset has been pre-filtered and formatted strictly for Vision training (e.g., DeepSeek-VL/OCR). It contains only the Free OCR task, with conversations mapped to the strict <|User|> and <|Assistant|> token format. Images are natively embedded.

---

Dataset Summary

Property	Value
Khasi Samples	28,668
English Samples	8,164
Languages	Khasi, English
Total Samples	36,832

---

Image Quality Distribution

The dataset intentionally includes both clean and degraded scans to improve OCR robustness.

Scan Quality	Samples
Clean Images	20,420
Degraded Images	16,412

---

Dataset Structure

Each sample contains the following fields:

• id — Unique identifier for the image.
• image — Document image embedded directly in the dataset.
• scan_quality — Image quality (clean or degraded).
• scan_type — Type of document (e.g., document, handwritten).
• language — Text language (kha for Khasi, en for English).
• messages — Instruction-style conversation used for Vision-Language training.

The conversation format follows DeepSeek-style roles:

• <|User|>
• <|Assistant|>

Example:

[
  {
    "role": "<|User|>",
    "content": "Free OCR.",
    "images": ["images/img_00048.jpg"]
  },
  {
    "role": "<|Assistant|>",
    "content": "Extracted markdown text from the document..."
  }
]

---

Tasks Included

Free OCR

The dataset focuses on transcribing document images into Markdown-formatted text.

The generated Markdown preserves basic document structure such as:

• headings
• paragraphs
• lists
• tables (when applicable)