GLINER detector
gliner-nvidia+homoglyph PII detection benchmark
Measured on 3 of 41 retained datasets under the same frozen scoring and masking protocol.
60.87%fully hidden annotations
67.05%detected / overlapped
2.86%extra masking
—CPU characters / second
Strengths
Highest complete-masking rates by normalized data type.
Phone numbers & email84.98%662 / 779
People's names76.01%1,131 / 1,488
Bank accounts & cards65.32%243 / 372
Watch list
Lowest observed complete-masking rates; inspect before deployment.
Passwords, keys & tokens42.56%163 / 383
Documents & identifiers47.67%1,280 / 2,685
Addresses & locations60.51%858 / 1,418
Language coverage
The overall leader can differ from the best choice for one language slice.
| Language | Datasets | Fully hidden ↑ | Detected ↑ | Extra masking ↓ |
|---|---|---|---|---|
| English | 1 | 29.86% | 47.57% | 3.29% |
| Russian | 2 | 62.11% | 67.82% | 1.90% |
Configuration, source and version
- Family
- GLINER
- Revision
- bd23e8ef4425fd04e34c5204ab49ffaa706eae79
- Eligible datasets
- 3 / 41
- Character F1
- 0.579
- Untouched diagnostic
- 32.95% · 2,467 / 7,486
- Benchmark version
- 1.0.3 · 2026-09-09
- CPU seconds / 10k chars
- Not measured
- GPU seconds / 10k chars
- 0.360
source overlap: nemotron-pii
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 | 29.86% | 47.57% | 3.29% |
| corrupt-redmadrobot | 5,531 | 60.12% | 64.36% | 1.25% |
| corrupt-hivetrace | 1,667 | 68.75% | 79.30% | 6.22% |