README.md

metadata

license: cdla-permissive-2.0
language:
  - en
tags:
  - benchmark
  - document-comprehension
  - office-documents
  - question-answering
  - llm-evaluation
pretty_name: OfficeComprehensionBenchmark (OCB)
size_categories:
  - 1K<n<10K
configs:
  - config_name: ocb_qna_data
    data_files:
      - split: train
        path: data/ocb_qna_data.parquet
  - config_name: ocb_source_urls
    data_files:
      - split: train
        path: data/ocb_source_urls.parquet

OfficeComprehensionBenchmark (OCB)

OCB is a benchmark for evaluating document comprehension and grounded reasoning over Microsoft Office files (Word, Excel, PowerPoint). It comprises two tracks:

• File Fidelity Q&A — measures structural and visual perception of document artifacts (text, tables, charts, formulas, formatting, embedded objects).
• Domain Q&A — measures expert-level reasoning over real-world business documents across 12 industries.

Companion repository

Evaluation prompts, judge scripts, file-download utilities, and PDF→Office conversion scripts are maintained in the companion GitHub repository:

https://github.com/microsoft/OfficeComprehensionBench

This Hugging Face dataset hosts the questions, reference answers, atomic assertion rubrics, the URL manifest for externally-hosted reference files, and the redistributable subset of reference files. Most reproduction tasks (running judges, fetching URL-referenced files, converting source PDFs to native Office formats) require the GitHub repository.

Benchmark size

• File Fidelity track: 244 files, 922 queries
• Domain Q&A track: 124 files, 120 queries, 8,450 atomic assertions across 12 industries

These numbers reflect the full evaluation set reported in the accompanying paper. See the May 07 Update note below for the relationship between the evaluated set and the publicly released artifact.

Distribution model

OCB uses a hybrid distribution model driven by source licensing:

• Hosted directly on Hugging Face — reference files whose source license permits redistribution under CDLA-Permissive-2.0, plus all questions, reference answers, and assertion rubrics.
• Referenced via URL manifest — reference files whose source licensing does not permit redistribution. The manifest is shipped with this dataset; the GitHub repository contains a release script that downloads each file from its original public source and, where required, converts it from its source format (e.g., PDF) into the native Office format (.docx, .xlsx, .pptx) used by the benchmark.

May 07 Update

Three Excel files originally included in the evaluation set are excluded from this public release for licensing reasons:

• winemag-data-130k-v2_sampled.csv
• Banking_Call_Data.csv
• brazilian_ecommerce_cleaned.csv

These files are sourced from Kaggle datasets published under CC BY-NC-SA 4.0, which is incompatible with the CDLA-Permissive-2.0 release license. The associated File Fidelity queries that target these files are also excluded from the public release.

Reported benchmark numbers in the paper reflect the full evaluation set including these files. Users wishing to reproduce the full evaluation can obtain the three source datasets directly from Kaggle under the original publishers' terms; URLs are listed in the URL manifest alongside the other externally-referenced files.

License

This dataset is released under the Community Data License Agreement – Permissive, Version 2.0 (CDLA-Permissive-2.0) — https://cdla.dev/permissive-2-0/.

Sub-components carry the following licenses:

• OCB-authored content (questions, reference answers, atomic assertions, expert-authored reference files): CDLA-Permissive-2.0.
• Hosted reference files from U.S. government open-data portals: U.S. Government Work / Public Domain (federal sources) or open-data terms permitting redistribution (state/local sources). Redistributed under CDLA-Permissive-2.0.
• Hosted reference files from Kaggle: only files under CC0 1.0 or MIT are redistributed. Attribution and notice obligations of MIT-licensed sources are preserved in the NOTICES.md file in this repository.
• URL-referenced files: retain their original licenses; OCB does not assert a license over these files. Users fetch them directly from the original publishers under those publishers' terms.

See NOTICES.md in this repository for per-source attribution and notice details.

Intended use

OCB is intended for evaluation, not training. Validated use cases include:

• Benchmarking model accuracy on grounded question answering over structured and semi-structured Office files
• Comparing models on cross-format reasoning (e.g., Word + Excel inputs)
• Evaluating fidelity of structural and visual document perception
• Ablation studies on retrieval and grounding components for Office-document pipelines

Not validated, and not recommended:

• Use as training data
• Evaluation on file formats outside .docx / .xlsx / .pptx and its variants
• High-stakes deployment decisions in regulated domains (medical, legal, financial advice) on the basis of OCB scores alone

Citation

@misc{shaik2026ocb,
  title  = {Office Comprehension Benchmark},
  author = {Firoz Shaik and Mateus Pican\c{c}o Lima Gomes and Tanvir Aumi
            and Jingci Wang and Milos Milunovic and Filip Basara
            and Ivana Jovanovic and Vishwas Suryanarayanan
            and Neha Nandan Kenkare and Weiyao Xie and Zhipeng Han
            and Zheng Zhang and Waleed Shahid and Jay Rathi
            and Russell Scherer and Thong Q. Nguyen and Michael Bentley
            and Tamara Stankovic and Rasika Chakravarthy and Vishal Chowdhary},
  year   = {2026}
}

Maintenance

Versioned releases are maintained on Hugging Face under this repository. Issues, corrections, and erratum requests are tracked through the dataset repository's discussion page. We commit to maintaining the dataset for at least 3 years post-publication.

If an externally-hosted file becomes unavailable from its original source, we will either replace it with an equivalent example or document the removal in the errata.