README.md

metadata

pretty_name: MedDocBench Sample
language:
  - en
  - zh
license: other
task_categories:
  - image-to-text
tags:
  - medical
  - ocr
  - document-understanding
  - clinical-forms
  - handwritten-text-recognition
  - vision-language-models
size_categories:
  - n<1K

MedDocBench Sample

This is the public 10-document sample release for MedDocBench: Benchmarking Vision-Language Models for Reliable Structured Extraction from Real-World Clinical Forms.

The sample is provided for repository smoke tests, format inspection, and reproducibility checks. The full secondary-redacted dataset is released as a separate gated Hugging Face dataset; its dataset card is public so users can read the access terms, but full-data downloads require manual approval.

Contents

data/
  images/                # 10 secondary-redacted PNG images
  sample_gt_v1.json       # 5 transcription-task annotations
  sample_gt_v2.json       # 5 structured-extraction annotations
  sample_gt.json          # legacy alias for v1
  dataset_split.json      # sample file lists for v1 and v2
metadata.jsonl            # lightweight index for inspection

Data Use

Use this sample for non-commercial research, education, evaluation, and reproducibility only. Do not attempt re-identification.

Contacts

Questions about MedDocBench or full dataset access should be directed to the corresponding authors:

• Weifa Yang: teddyrun@hku.hk
• Lequan Yu: lqyu@hku.hk