Skip to content

Dataset · jiangxuhao

ExamOCRBench schoolwork and handwriting recognition benchmark

jiangxuhao/ExamOCRBench

Annotated dataset with photos of early school exam pages for evaluating multimodal OCR and VQA.

What's inside

The public annotated part of the ExamOCRBench benchmark includes 24,928 examples in Chinese and English. Inside are tasks for recognizing photographed school exam pages, handwritten answers, document structure, and visual grounding. Ground truth for all four tasks is stored in the ground_truth_json format.

What you can do with it

  • Evaluating multimodal large language model (VLM) performance in handwritten text recognition (OCR) tasks
  • Testing visual question answering (VQA) systems on complex educational documents
  • Fine-tuning element detection and structure parsing models for scanned pages
Tasks
Image → text, Visual question answering, Object detection
Languages
Chinese, English
Size
10K–100K rows
Downloads this month
2.5K
Likes
9

How to load it

from datasets import load_dataset

ds = load_dataset("jiangxuhao/ExamOCRBench")