name: ArenaOCR
description: ArenaOCR is an extremely challenging document OCR benchmark featuring 5,000 unique, procedurally generated PDF pages and unit test cases. It covers six core categories of complex OCR challenges including academic arXiv mathematics, dense and multi-column tables, complex header/footer metadata, multi-column reading orders, degraded photocopy-style old scans, and dense microscopic text.
evaluation_framework: arenaocr

tasks:
  - id: ArenaOCR_ArXivMath
    config: default
    split: arxiv_math
  - id: ArenaOCR_HeadersFooters
    config: default
    split: headers_footers
  - id: ArenaOCR_TableTests
    config: default
    split: table_tests
  - id: ArenaOCR_MultiColumn
    config: default
    split: multi_column
  - id: ArenaOCR_OldScans
    config: default
    split: old_scans
  - id: ArenaOCR_LongTinyText
    config: default
    split: long_tiny_text