ArenaOCR Benchmark

ArenaOCR is a highly rigorous, unit-test-driven Optical Character Recognition (OCR) and Document Understanding benchmark designed to assess the performance of Vision-Language Models (VLMs) and advanced OCR systems on extremely challenging real-world layouts.

Replicating the design paradigm and schema structure of allenai/olmOCR-bench, ArenaOCR shifts away from traditional "fuzzy" metrics (like character error rate, edit distance, or BLEU/ROUGE) and instead evaluates document transcripts using machine-verifiable, deterministic unit tests (e.g. math formula accuracy, column order preservation, header/footer suppression, and noise-tolerant transcription).

---

Dataset Splits & Tasks

ArenaOCR contains 5,000 unique, procedurally generated PDF documents and their corresponding JSONL unit tests split across 6 key difficulty divisions:

1. arxiv_math (850 samples): Evaluation of complex, multi-level academic LaTeX mathematical equations, featuring nested fractions, integrals, sums, Greek characters, and matrices.
2. headers_footers (830 samples): Assesses whether OCR systems can successfully isolate the document's central body text while discarding page-margin metadata like running headers, page counts, and publication tags.
3. table_tests (830 samples): Complex multi-column/multi-row layouts featuring cell merges (SPAN), missing cell boundaries, alternating shading, and dense finance/science alphanumeric matrices.
4. multi_column (830 samples): 2-column or 3-column academic article structures. Evaluates reading order preservation, verifying that the OCR reads columns vertically rather than leaking text horizontally across separators.
5. old_scans (830 samples): Simulates degraded photocopy text sheets from vintage manuscripts, featuring random speckle noise, page skew, faded inks, and streaking lines.
6. long_tiny_text (830 samples): Exceedingly dense legal terms and conditions (TOS/NDA agreements) utilizing minuscule (4.5pt - 5.5pt) font sizes to test transcription precision.

---

Dataset Schema

Each JSONL unit test entry contains:

• pdf_filename (string): Relative path to the PDF file (e.g., bench_data/pdfs/arxiv_math/arxiv_math_0001.pdf).
• page_number (int): Page number within the document (always 1 for single-page benchmark pages).
• test_type (string): The verification logic applied:
  • math_formula: LaTeX comparison of mathematical expressions.
  • text_absence: Verifies that margins or header information were excluded.
  • text_presence: Substring search validating target text extraction.
  • reading_order: Checks if first_text occurs in the transcript before second_text.
• text (string, optional): String parameter for presence/absence checks.
• case_sensitive (bool, optional): Determines case matching constraints for presence/absence.
• formula (string, optional): Exact LaTeX ground-truth target.
• first_text (string, optional): Anchoring phrase that must appear earlier.
• second_text (string, optional): Anchoring phrase that must appear later.

---

Local Evaluation

A local evaluation script eval_bench.py is included in the repository. Running the following command will evaluate model transcripts saved in a ./predictions directory against our benchmark unit tests:

python eval_bench.py --predictions ./predictions