Skip to content

Dataset · zhihengli-casia

Three-mode document recognition benchmark

zhihengli-casia/puredocbench

From 1 to 10k document pages for comparing OCR and vision-language models across three tracks.

What's inside

A dataset and leaderboard for evaluating document recognition (image-to-text) and document question-answering, ranging from 1,000 to 10,000 entries. Evaluation is split into three tracks with 1,475 pages each: Clean, Digital, and Real documents. The card presents results for 58 models, including pipeline solutions, specialized end-to-end models, and general VLMs.

What you can do with it

  • Comparing the quality of OCR systems and VLMs on clean, digital, and real documents
  • Evaluating models in text extraction and document page question-answering tasks
Tasks
Image → text, Document question answering
Size
1K–10K rows
Downloads this month
3.4K
Likes
6

How to load it

from datasets import load_dataset

ds = load_dataset("zhihengli-casia/puredocbench")