GLINER2 detector
gliner25-fastino-ru+nochunk PII detection benchmark
Measured on 5 of 41 retained datasets under the same frozen scoring and masking protocol.
45.53%fully hidden annotations
48.19%detected / overlapped
1.59%extra masking
—CPU characters / second
Strengths
Highest complete-masking rates by normalized data type.
Phone numbers & email71.92%3,127 / 4,348
Passwords, keys & tokens70.13%263 / 375
Bank accounts & cards66.15%43 / 65
Watch list
Lowest observed complete-masking rates; inspect before deployment.
Logins & usernames12.69%960 / 7,565
Network identifiers19.43%3,250 / 16,729
Customer & employee IDs27.97%547 / 1,956
Language coverage
The overall leader can differ from the best choice for one language slice.
| Language | Datasets | Fully hidden ↑ | Detected ↑ | Extra masking ↓ |
|---|---|---|---|---|
| Russian | 5 | 45.53% | 48.19% | 1.59% |
Configuration, source and version
- Family
- GLINER2
- Revision
- aaecfe45db1d828c963717054ccb868e8ad1f1d5
- Eligible datasets
- 5 / 41
- Character F1
- 0.577
- Untouched diagnostic
- 51.81% · 38,785 / 74,859
- Benchmark version
- 1.0.3 · 2026-09-09
- CPU seconds / 10k chars
- Not measured
- GPU seconds / 10k chars
- Not measured
Russian labels
Results by dataset
Complete masking is primary; detection shows whether any character overlapped.
| Dataset | Gold annotations | Fully hidden ↑ | Detected ↑ | Extra masking ↓ |
|---|---|---|---|---|
| synth-ru-tickets | 32,099 | 27.76% | 29.42% | — |
| factrueval | 7,966 | 50.74% | 51.51% | — |
| redact-ru | 9,230 | 59.34% | 63.20% | 11.14% |
| nerel | 24,369 | 60.66% | 64.31% | — |
| jayguard | 1,195 | 72.97% | 85.69% | 1.35% |