Local, reversible PII anonymization for Italian legal text.
Desktop app (no setup, runs offline)
- Rizzo-PII-1.0.0-Windows-Setup.exe — Windows installer (per-user, no admin needed).
- Rizzo-PII-1.0.0-Linux-x86_64.AppImage — Linux portable build (
chmod +xthen run). - Rizzo-PII-1.0.0-macOS-arm64.dmg — macOS build (Apple Silicon / arm64). Signed & notarized by Apple: open the
.dmg, drag the app to Applications, and launch it — no Gatekeeper warnings.
They run 100% offline on CPU — no GPU, no API key, no telemetry. They detect 22 PII categories incl. Italian-legal CF / PIVA / dati catastali, paired with a regex+checksum safety net, and drive a reversible workflow: anonymize locally, query any frontier LLM with placeholders, restore locally.
Model weights (for developers)
- rizzo-pii-0.3B-v1.2.0.zip — the trained mmBERT token-classifier (micro-F1 0.989 on real Italian validation).
Unzip it into the repo's models/ folder; the scripts auto-detect the latest rizzo-pii-0.3B-v*, or set PII_MODEL_DIR to the extracted path:
python src/training/test_pii.py "Mi chiamo Mario Rossi, IBAN IT60X0542811101000000123456"