Oriel — AI & technology worth noticing, every day.

baidu/MarathonDocBench

Dataset · https://huggingface.co/datasets/baidu/MarathonDocBench

MarathonDocBench A New Benchmark for Ultra-Long Document Parsing: Targeting Hundred-Page-Scale Documents. MarathonDocBench is the first benchmark specifically designed for hundred-page document parsing. It contains 100 documents spanning 100–300 pages (~18k pages, ~205k annotated layout elements, mixed Chinese/English), with rich cross-page structures — cross-page text and tables appear frequently and densely across documents.… See the full description on the dataset page: https://huggingface.co/datasets/baidu/MarathonDocBench.

First seen by Oriel on .

AI summary

MarathonDocBench is a benchmark for parsing very long documents, providing 100 annotated Chinese and English documents of 100–300 pages and an eight-metric evaluation setup.

  • The dataset includes about 18,000 page images and 205,000 annotated layout elements, including reading order, hierarchy, bounding boxes, and cross-page text and table continuations.
  • Annotations start from an automatic page-level parser and are refined by human annotators; transcription disagreements flagged by a VLM pipeline receive human adjudication.
  • Evaluation uses the MPDocBench-Parse schema and toolkit, with supplied patches for parallel processing and configurable timeouts; a full run takes tens of minutes with the stated settings.
  • The leaderboard reports Unlimited OCR Works 2 at 77.06 Overall, where Overall is the mean of eight metrics; long inputs can exceed provider limits or truncate outputs.
  • Annotations and tooling references use Apache-2.0, while page images are for academic and research use and the license does not grant redistribution of source documents.

#document parsing #ocr #benchmark #multimodal models

Generated by gpt-5.6-sol. Check the original for details.

Metric trends

Last 30 days · Up to 120 recorded samples per source and metric, with reposts kept separate · First and latest retained

Likes

Total likes reported by the source.

  • Hugging Face Datasets: 20, latest observation ; +9 over observed period since

Downloads (30d)

Rolling 30-day downloads. A change in this value is not new downloads during the observed period.

  • Hugging Face Datasets: 5,706, latest observation ; +5,688 over observed period since

Aliases

SourceExternal IDURL
Hugging Face Datasetsbaidu/MarathonDocBenchhttps://huggingface.co/datasets/baidu/MarathonDocBench Read discussion