PHI Mask

Detection benchmarks.

The Box Draw Eval uses authored static documents and interfaces. Explore its source gallery and current detector coverage in Box Draw Eval gallery.

Browser payload

  • 0.6 MBtext
  • 0.4 MBPDF text layer
  • 4.1 MBOCR images & scanned PDFs, English

Fully works offline in Airplane Mode

Updated August 26, 2026

PHI Mask results by capability, grouped by document type
CapabilityPHI Mask result
Box Draw Eval source corpus
Redaction-footprint standard · v1Detection Gold is the tight visible footprint for one value the detector should recognize. It uses the rendered word box for one word or the combined contiguous word boxes for several words.Last-measured percentages report Detection Gold targets found under the existing lane rules. Ownership and mask or keep decisions are scored separately.
100.0%
Healthcare13 mocks · 594/594 detection targets found100.0%
Hospital & clinic EHR7 mocks · 300/300 detection targets found100.0%
Regional health records3 mocks · 225/225 detection targets found100.0%
Pharmacy1 mocks · 27/27 detection targets found100.0%
Clinical data on everyday tools2 mocks · 42/42 detection targets found100.0%
Finance1 mocks · 6/6 detection targets found100.0%
Banking & investment1 mocks · 6/6 detection targets found100.0%
Structured documentsInvoices, receipts and statements — sensitive values sit in labelled fields.
Sensitive values blacked outFATURA annotation corpus100%
Sensitive values blacked out — scanned imagesFATURA scanned-image corpus97.6%
Nothing left partly visibleFATURA annotation corpus100.0%
Names and account numbers blacked outInvoiceBenchmark corpus100%
Nothing blacked out by mistake — invoicesInvoiceBenchmark corpus100%
Nothing blacked out by mistake — receiptsCORD v2 corpus100%
Loose sentencesNotes, messages and prose — nothing marks where the sensitive part is.
People's namesPresidio synthetic corpus100.0%
Identity tokens blacked out correctlyPresidio synthetic corpus96.8%
Census-derived namesOfficial population-name aggregates projected into synthetic pairs — no registry rows, business rows or addresses.
Name targets detectedCensus-derived synthetic name corpus100.0%
Clinical text coverageMEDDOCAN is shown as retained entity coverage only. Character leakage was not retained, and no overall score is claimed.
Retained clinical entity coverageMEDDOCAN retained replay · character leakage unavailable96.7%

Source notices and attribution

These rows measure PHI Mask on upstream datasets or deterministic synthetic projections. Upstream source terms and licenses remain authoritative; public PHI Mask artifacts publish aggregate results only and do not redistribute source rows, images, annotations, or recovered values.

  • FATURA: FATURA was created by Mahmoud Limam, Marwa Dhiaf, and Yousri Kessentini at the Digital Research Center of Sfax. PHI Mask publishes aggregate annotation and OCR measurements only; source images and annotations are not redistributed. Zenodo record, paper.
  • Presidio synth_dataset: The legacy synthetic corpus and evaluator are from Smile-SA's anonymization project; the pinned source revision is recorded in the public aggregate. PHI Mask publishes counts and hashes only. source corpus, evaluator.
  • InvoiceBenchmark: InvoiceBenchmark is attributed to jngb-labs; its Hugging Face dataset page identifies the dataset as MIT-licensed. PHI Mask publishes a derived masking measurement with aggregate counts only. dataset card and license.
  • CORD v2: CORD: A Consolidated Receipt Dataset for Post-OCR Parsing is attributed to Park, Shin, Lee, Surh, Seo, and the CORD authors; the source repository identifies the dataset as CC BY 4.0. PHI Mask uses it only as a hard-negative characterization and publishes aggregate counts. source and citation, CC BY 4.0 license.
  • Census-derived names: The census row uses aggregate name statistics from the U.S. Census Bureau, Statistics Canada, IBGE, and Statistics Greenland. PHI Mask publishes deterministic synthetic pairings and observed shape counts only; no source values or reconstructed names are published. U.S. Census Bureau, Statistics Canada, IBGE, Statistics Greenland NAXT3, Statistics Greenland NAXT1.
  • MEDDOCAN / SPACCC_MEDDOCAN: This work is licensed under the Creative Commons Attribution 4.0 International License. Copyright (c) 2019 Secretaría de Estado para el Avance Digital (SEAD). Changes made: PHI Mask publishes an aggregate retained replay only; source rows, values, annotations, and row-level output are not included. source project, Creative Commons Attribution 4.0 International License.