PPLX detector
pplx+sent300 PII detection benchmark
Measured on 11 of 41 retained datasets under the same frozen scoring and masking protocol.
65.39%fully hidden annotations
68.78%detected / overlapped
8.97%extra masking
—CPU characters / second
Strengths
Highest complete-masking rates by normalized data type.
Logins & usernames100.00%7,794 / 7,794
Network identifiers98.30%17,201 / 17,498
Phone numbers & email97.02%5,576 / 5,747
Watch list
Lowest observed complete-masking rates; inspect before deployment.
Organizations4.11%540 / 13,123
Addresses & locations26.33%4,038 / 15,334
Other sensitive attributes37.57%334 / 889
Language coverage
The overall leader can differ from the best choice for one language slice.
| Language | Datasets | Fully hidden ↑ | Detected ↑ | Extra masking ↓ |
|---|---|---|---|---|
| English | 6 | 70.12% | 77.65% | 9.95% |
| Russian | 5 | 64.28% | 66.70% | 2.52% |
Configuration, source and version
- Family
- PPLX
- Revision
- f1f90a53823f5df0a1344c1e137d9fffdaab54d6
- Eligible datasets
- 11 / 41
- Character F1
- 0.687
- Untouched diagnostic
- 31.22% · 28,830 / 92,332
- Benchmark version
- 1.0.3 · 2026-09-09
- CPU seconds / 10k chars
- Not measured
- GPU seconds / 10k chars
- 2.837
Results by dataset
Complete masking is primary; detection shows whether any character overlapped.
| Dataset | Gold annotations | Fully hidden ↑ | Detected ↑ | Extra masking ↓ |
|---|---|---|---|---|
| factrueval | 7,966 | 27.40% | 27.49% | — |
| nerel | 24,369 | 27.73% | 29.06% | — |
| tab-echr | 3,830 | 28.41% | 47.36% | — |
| dialogpii-en | 3,087 | 69.97% | 75.83% | 5.37% |
| jayguard | 1,195 | 72.97% | 84.69% | 2.37% |
| secrets-issues | 288 | 73.61% | 96.88% | 10.68% |
| leak-museum | 101 | 82.18% | 83.17% | 3.45% |
| redact-ru | 9,230 | 84.31% | 90.02% | 8.67% |
| nemotron-pii | 9,391 | 85.65% | 88.62% | 7.23% |
| creddata | 776 | 85.82% | 93.81% | 8.63% |
| synth-ru-tickets | 32,099 | 95.11% | 97.64% | — |