README.md

metadata

language:
  - th
license: apache-2.0
task_categories:
  - image-to-text
tags:
  - ocr
  - thai
  - synthetic
size_categories:
  - 1K<n<10K

Thai OCR Dataset

Synthetic Thai OCR dataset for fine-tuning vision-language models (e.g., DeepSeek-OCR-2).

Dataset Summary

• Total samples: 4000
• Subsets:
  • text — General Thai document text (1,000 samples)
  • table — Thai table formats: invoices, budgets, schedules, etc. (1,000 samples)
  • official — Thai government documents: contracts, legal, police reports, official letters (2,000 samples)

Format

Parquet with embedded images. Compatible with HF dataset viewer.

Columns: image, text, subset

Usage

from datasets import load_dataset

ds = load_dataset("mekpro/ocr_th")

# Filter by subset
tables = ds.filter(lambda x: x["subset"] == "table")
official = ds.filter(lambda x: x["subset"] == "official")