README.md

metadata

license: cc-by-4.0
task_categories:
  - image-to-text
language:
  - zh
  - en
tags:
  - ocr
  - vision
  - multimodal
  - document-understanding
  - image-text-to-text
size_categories:
  - 10K<n<100K
pretty_name: TextRenderLab OCR Dataset

TextRenderLab OCR Dataset

This dataset is a synthetic OCR instruction-tuning dataset built with TextRenderLab. It contains rendered text images paired with OCR prompts and full transcription targets in both Traditional Chinese and English.

Dataset Summary

• Task: image-to-text OCR transcription
• Languages: Traditional Chinese, English
• Total samples: 40,000
• Language split:
  • Chinese: 20,000
  • English: 20,000
• Image format: PNG
• Annotation format: JSONL

Each sample contains:

1. A rendered image path under images/
2. An OCR-style user question in Chinese
3. A full text transcription as the answer
4. Chat-style fields for multimodal training pipelines

Data Structure

Each line in dataset.jsonl is a JSON object with the following fields:

Field	Type	Description
unique_id	string	Unique sample identifier
images	list[string]	Relative image path list, typically one PNG file
question	string	OCR prompt
answer	string	Full transcription target
messages	list[object]	Chat-style training format with <image> prompt
axolotl_messages	list[object]	Axolotl-compatible multimodal message format

Example

{
  "unique_id": "0ff0e0d5-afcf-48aa-a08b-c6d1a1ddc5eb",
  "images": ["images/en_ocr_00000.png"],
  "question": "圖片裡有些字，請幫我轉錄出來。",
  "answer": "Under the legal framework that defined Hong Kong’s semi-autonomy after its return to China in 1997, many in the territory assumed that it would mostly manage its own affairs for decades.",
  "messages": [
    {"role": "user", "content": "<image>圖片裡有些字，請幫我轉錄出來。"},
    {"role": "assistant", "content": "Under the legal framework that defined Hong Kong’s semi-autonomy after its return to China in 1997, many in the territory assumed that it would mostly manage its own affairs for decades."}
  ],
  "axolotl_messages": [
    {
      "role": "user",
      "content": [
        {"type": "image", "url": "images/en_ocr_00000.png"},
        {"type": "text", "text": "圖片裡有些字，請幫我轉錄出來。"}
      ]
    },
    {
      "role": "assistant",
      "content": [
        {"type": "text", "text": "Under the legal framework that defined Hong Kong’s semi-autonomy after its return to China in 1997, many in the territory assumed that it would mostly manage its own affairs for decades."}
      ]
    }
  ]
}

Intended Use

This dataset is intended for:

• OCR fine-tuning for multimodal LLMs
• Image-text transcription training
• Bilingual OCR robustness experiments
• Synthetic pretraining or instruction-tuning pipelines

Notes

• The images are synthetically rendered rather than photographed.
• Prompts are phrased in Chinese, while answers may be either Chinese or English depending on the source text.
• Relative image paths in images assume the repository or dataset root contains both dataset.jsonl and the images/ directory.

Loading Example

import json
from pathlib import Path

dataset_path = Path("dataset.jsonl")
with dataset_path.open("r", encoding="utf-8") as f:
    first = json.loads(next(f))

print(first["images"])
print(first["question"])
print(first["answer"][:200])

Citation

If you use this dataset, please cite or reference the Hugging Face dataset repository:

minyichen/Render-OCR