Measuring what matters: not just whether PII is found, but whether your detector understands context.
F1 scores across all four benchmark categories. Higher is better.
| System | Overall F1 | Standard | Ambiguous | CSS | Adversarial |
|---|
Can your detector classify the same string differently based on context?
"Jordan presented the quarterly results to the board."
PERSON"The shipment arrived in Jordan yesterday."
LOCATIONSame string, different entity types. A system that classifies both correctly scores +1. Both wrong = 0.
How each system performs on individual entity types.
Radar chart showing each system's strengths across categories.
Per-document server-side processing time. AmbientMeta times reflect actual detection processing, not network round-trip.
| System | p50 (ms) | p95 (ms) | p99 (ms) | Throughput |
|---|
Clone, install, run. Reproducible results in under 5 minutes.
git clone https://github.com/ambientmeta/ambientmeta-pii-detection-benchmark
cd ambientmeta-pii-detection-benchmark
pip install -e .
python -m spacy download en_core_web_lg
python run_benchmark.py
The benchmark runs without an AmbientMeta API key. Set AMBIENTMETA_API_KEY to include AmbientMeta in the comparison.