HF detector
ru-legal-ner+homoglyph PII detection benchmark
Measured on 3 of 41 retained datasets under the same frozen scoring and masking protocol.
59.24%fully hidden annotations
76.58%detected / overlapped
11.07%extra masking
—CPU characters / second
Strengths
Highest complete-masking rates by normalized data type.
People's names74.40%1,107 / 1,488
Documents & identifiers70.02%1,880 / 2,685
Phone numbers & email64.57%503 / 779
Watch list
Lowest observed complete-masking rates; inspect before deployment.
Network identifiers3.05%11 / 361
Addresses & locations38.01%539 / 1,418
Passwords, keys & tokens49.35%189 / 383
Language coverage
The overall leader can differ from the best choice for one language slice.
| Language | Datasets | Fully hidden ↑ | Detected ↑ | Extra masking ↓ |
|---|---|---|---|---|
| English | 1 | 41.32% | 77.08% | 10.02% |
| Russian | 2 | 59.96% | 76.56% | 13.46% |
Configuration, source and version
- Family
- HF
- Revision
- c0313a42ac147ccc38f5c3b0b3e77cab53208683
- Eligible datasets
- 3 / 41
- Character F1
- 0.387
- Untouched diagnostic
- 23.42% · 1,753 / 7,486
- Benchmark version
- 1.0.3 · 2026-09-09
- CPU seconds / 10k chars
- Not measured
- GPU seconds / 10k chars
- 0.143
Results by dataset
Complete masking is primary; detection shows whether any character overlapped.
| Dataset | Gold annotations | Fully hidden ↑ | Detected ↑ | Extra masking ↓ |
|---|---|---|---|---|
| corrupt-secrets-issues | 288 | 41.32% | 77.08% | 10.02% |
| corrupt-redmadrobot | 5,531 | 57.35% | 73.02% | 12.69% |
| corrupt-hivetrace | 1,667 | 68.63% | 88.30% | 18.65% |