Detection Benchmarks
Security claims should be inspectable. We continuously test our open-source detection engine against standardized datasets to measure accuracy, precision, and recall.
Pseudonymization reduces exposure. It does not make data anonymous. For a privacy product, recall matters enormously because false negatives leak data to external APIs.
Current Benchmark Results
- Engine Version: 1.0.0
- Commit SHA:
f8387dc - Date: September 6, 2026
- Dataset: Enron / Presidio standard sets (Synthetic & Real)
- Sample size: 10,000 sentences
- Languages: English, Italian
| Entity | Precision | Recall | F1 Score |
|---|---|---|---|
| 0.99 | 0.99 | 0.99 | |
| Person | 0.94 | 0.88 | 0.91 |
| Location | 0.91 | 0.85 | 0.88 |
| Organization | 0.90 | 0.82 | 0.86 |
| Phone | 0.98 | 0.95 | 0.96 |
| IP Address | 1.00 | 1.00 | 1.00 |
Methodology
- Precision: Out of all the entities the engine flagged as "Person", how many were actually people? (A low precision means high false positives, destroying useful non-sensitive text).
- Recall: Out of all the actual people in the text, how many did the engine successfully flag? (A low recall means high false negatives, leaking sensitive data).
- F1 Score: The harmonic mean of precision and recall.
Reproduction
You can reproduce these benchmarks yourself using our public Python package.
pip install "pseudonymize[ml]"
python -m pseudonymize.benchmarks --benchmark-only