Papers.Data.Code
Dataset

MarathonDocBench

baidu

Contains 100 ultra-long Chinese/English documents as page images with document-level layout annotations for OCR and parsing across many pages.

Hugging Face ↗Oct 26, 2026Code ↗

  • multimodal
  • cv
  • document-parsing
  • ocr
  • long-document
Comprehensive benchmark comparison showing method dominating baselines across all OCR metrics
Comprehensive benchmark comparison showing method dominating baselines across all OCR metrics

What it is

Benchmark for parsing hundred-page documents with page images and dense cross-page annotations. It targets ultra-long OCR/layout extraction where page-wise parsers break. 100 documents of 100-300 pages each (~18k pages, ~205k layout elements), with page images plus bilingual annotations for boxes, types, hierarchy, reading order, and cross-page text/table continuations.

Key results

  • 100 documents spanning 100-300 pages each
  • ~18k PNG pages and ~205k annotated layout elements
  • Annotations include boxes, order, hierarchy, and cross-page merges

Useful for benchmarking OCR and layout parsers on hundred-page documents with cross-page text and tables.

One useful ML email a week.

More from Oct 26 – Nov 01

The whole week →

Picked Oct 30

Picked Oct 29

  • Dataset
    crtproduction/name-that-llm-open-pool

    Unlike most attribution datasets, it provides 254,561 de-duplicated open passages labeled across 18 LLMs with provenance and leakage flags.

    KaggleOct 28, 2026
    →
  • Repo
    google-deepmind/detectorscope

    Unlike black-box AI text detectors, it exposes named feature sums and token-level evidence while preserving accuracy above 0.7 interpretability on COLING and RAID.

    GitHubOct 24, 2026
    →

Picked Oct 27

  • Dataset
    Qwen/BabelArena

    Unlike translation-only multilingual benchmarks, it fully localizes tasks, tools, and environments across 23 languages.

    Hugging FaceOct 22, 2026
    →

Worth your time, not all of your time.

We track new ML work across papers, code, and data, then keep only the pieces worth a closer look. Selected, not collected.

Papers.Data.Code

The Weekly ML Signal

Selected, not collected.

One useful ML email a week: the papers, repos and datasets worth your attention.

Worth your time, not all of your time.