How to read a personal data benchmark: recall, precision and false alarm
By Gurubase, .
Overview
Read three numbers together, because each one hides what the others show. Recall says how much personal data a system finds. Precision says how many of its masks were correct. False alarm says how often it masks clean text.
The numbers below come from our benchmark: 403 human-verified Turkish documents, October 2026.
Recall
Recall: how much personal data the system finds
Recall is the share of personal data that the system finds. Siper’s recall is 85.7%, and the next system’s is 69.3%.
Take a text with 100 pieces of personal data. At this recall, Siper finds about 86. The 95% confidence interval for Siper is 0.839 to 0.877.
Precision
Precision: how many masks were correct
Precision is the share of masks that hit real personal data. Siper’s precision is 56.1%. Higher precision means fewer unnecessary masks.
High precision alone proves little. One system in the table below has 91.8% precision and 26.7% recall. A system that masks little can make few wrong masks, so check recall next to it.
False alarm
False alarm: how often clean text gets masked
False alarm is the share of clean documents that receive at least one unnecessary mask. We measure it on 321 documents with no personal data. Siper’s rate is 14.3%.
Siper is tuned toward recall, so it finds more of the personal data.
Together
Read them side by side
The right balance depends on your work, so compare all three columns.
| System | Recall | Precision | False alarm |
|---|---|---|---|
| Gurubase Siper | 85.7% | 56.1% | 14.3% |
| GPT-5-mini | 69.3% | 62.8% | 12.1% |
| Gemini 2.5 Flash | 52.2% | 61.5% | 1.9% |
| Limina (Private AI) | 51.5% | 65.2% | 48.6% |
| Azure AI Language | 26.7% | 91.8% | 3.4% |
Special categories
Look at the sensitive categories
An overall average can hide the data you care about most, so check each category. On health data, Siper finds 81% and the next system finds 51%. On criminal convictions, Siper finds 77% and the next system finds 53%.
Your own text
Test it on your own text
Your documents are the test that counts. Keep your firewall and DLP. Test how much Turkish data they find on your own text.
Read the full benchmark, or try the playground.
Next step