PII Arena
ONNX detector

openai-base-onnx PII detection benchmark

Measured on 9 of 41 retained datasets under the same frozen scoring and masking protocol.

70.86%fully hidden annotations
76.66%detected / overlapped
3.20%extra masking
CPU characters / second

Strengths

Highest complete-masking rates by normalized data type.

Dates & times96.29%1,792 / 1,861
Phone numbers & email93.44%6,063 / 6,489
Bank accounts & cards86.17%1,620 / 1,880

Watch list

Lowest observed complete-masking rates; inspect before deployment.

Other sensitive attributes5.36%12 / 224
Organizations9.79%447 / 4,568
Addresses & locations53.79%4,791 / 8,907

Language coverage

The overall leader can differ from the best choice for one language slice.

LanguageDatasetsFully hidden ↑Detected ↑Extra masking ↓
English455.31%57.09%3.84%
Russian474.97%82.76%1.42%
Multilingual176.75%78.02%

Configuration, source and version

Family
ONNX
Revision
9b7e2f5c794da2c156bd9583bf8e2f5f6bdfd18b
Eligible datasets
9 / 41
Character F1
0.713
Untouched diagnostic
23.34% · 16,829 / 72,095
Benchmark version
1.0.3 · 2026-09-09
CPU seconds / 10k chars
Not measured
GPU seconds / 10k chars
Not measured

mirror of openai-base

Results by dataset

Complete masking is primary; detection shows whether any character overlapped.

DatasetGold annotationsFully hidden ↑Detected ↑Extra masking ↓
tab-echr3,83022.35%23.16%
ameau012,24253.30%54.86%2.58%
redmadrobot5,51665.19%70.43%1.39%
nemotron-pii9,39168.76%70.38%2.55%
secrets-issues28870.83%92.71%4.09%
synth-ru-tickets32,09973.82%83.05%
kiji-multi7,56576.75%78.02%
nym-ru9,49782.43%87.68%
hivetrace1,66786.98%90.10%1.66%