---

rizzo-pii-0.3B is a lightweight, CPU-friendly, Italian-first token-classification model (≈0.3B parameters, mmBERT / ModernBERT backbone) that detects 22 categories of personal data — including the Italian legal identifiers (codice fiscale, partita IVA, dati catastali) that no other open model covers — to drive a fully reversible anonymization workflow:

your document → [FULLNAME_1], [IBAN_1], [CF_1] … + local dictionary → closed LLM API → reconstruction

The sensitive values never leave your machine. Only the placeholders go to the API; the local dictionary rebuilds the original from the model's answer. Built for law firms and GDPR compliance.

---

🚀 Quick start

from transformers import pipeline

nlp = pipeline(
    "token-classification",
    model="rizzoaiacademy/rizzo-pii-0.3B",
    aggregation_strategy="simple",   # merges B-/I- subwords into whole entities
)

text = ("Mi chiamo Mario Rossi, codice fiscale RSSMRA85M01H501Z, "
        "IBAN IT60X0542811101000000123456, email mario.rossi@gmail.com.")

for ent in nlp(text):
    print(f"{ent['entity_group']:<14} {ent['word']!r}  ({ent['score']:.2f})")

# FULLNAME       'Mario Rossi'                   (1.00)
# CF             'RSSMRA85M01H501Z'              (1.00)
# IBAN           'IT60X0542811101000000123456'  (1.00)
# EMAIL          'mario.rossi@gmail.com'         (1.00)

🔁 Reversible anonymization (the intended use)

def anonymize(text, ents):
    """Replace each entity with a numbered placeholder; keep a reversible dictionary."""
    mapping, out, counters = {}, text, {}
    for e in sorted(ents, key=lambda x: x["start"], reverse=True):
        g = e["entity_group"]
        counters[g] = counters.get(g, 0) + 1
        tag = f"[{g}_{counters[g]}]"
        mapping[tag] = text[e["start"]:e["end"]]
        out = out[:e["start"]] + tag + out[e["end"]:]
    return out, mapping

anon, table = anonymize(text, nlp(text))
# anon  -> "Mi chiamo [FULLNAME_1], codice fiscale [CF_1], IBAN [IBAN_1], email [EMAIL_1]."
# table -> {"[CF_1]": "RSSMRA85M01H501Z", ...}   # stays local; use it to rebuild the LLM reply

💡 Production tip: always pair the model with a regex + checksum safety net for structured fields (EMAIL / TELEPHONE / IBAN / CF / PIVA / credit-card / amount / plate). IBAN, CF, PIVA and card numbers are mathematically verifiable — let the checksum override the model when they disagree. This is exactly what the rizzo-pii desktop app does.

---

🏷️ The 22 PII categories

Labels use the BIO scheme → 44 label ids (B-/I- per tag + O). Full taxonomy and merge rules in the project docs.

---

📊 Evaluation

Entity-level metrics on a held-out Italian validation set (7,000 real sentences; the 5 IT-legal tags injected into held-out real text — real context, no leakage):

Validation is Italian-only by design (the real use case is the IT legal domain). The training is multilingual, but the other 7 languages are not validated here.

---

🧠 Training

• Base model: jhu-clsp/mmBERT-base (multilingual encoder, ModernBERT architecture, native 8192 context). Chosen over vanilla ModernBERT because the latter is almost English-only.
• Data: ≈745k rows fused from 4 sources (multilingual; Italian reinforced to ~45%, ~38% synthetic). See the dataset: 🤗 rizzoaiacademy/rizzo-pii-it-dataset.
• Recipe: 1 epoch, max_len=768, effective batch 32, dynamic padding, group_by_length.
• Hardware: single RTX 5060 Ti (Blackwell, 16 GB).

Key idea: "LLM as author, code as labeler"

Synthetic data is generated by letting an LLM write only the prose with placeholders, while code injects the values. This gives exact BIO labels, mathematically valid checksums (CF/PIVA/IBAN), and guarantees no real PII is ever produced by the LLM.

---

⚠️ Limitations & honest expectations

• Structural overfit risk on tags that come only from templates (CATASTO, PROVINCE) — mitigated with 72 templates, real-text augmentation, and DeepMount real context.
• Italian-only validation — the 7 non-IT languages are not measured.
• Class imbalance — FULLNAME ≫ CREDITCARDNUMBER (~66×), so rare tags are noisier.
• Always add a regex+checksum net in production. The model is the recall engine; the checksum is the precision guarantee for structured identifiers.

---

🔗 Links

• 💻 App + code (GitHub): https://github.com/Rizzo-AI-Academy/rizzo-pii — Windows installer, Flask/Tauri UI
• 📚 Training dataset: https://huggingface.co/datasets/rizzoaiacademy/rizzo-pii-it-dataset
• 📄 Full technical report (PDF) in the GitHub repo

📄 License

MIT © 2026 Simone Rizzo — Rizzo AI Academy

📌 Citation

@software{rizzo_pii_2026,
  author  = {Simone Rizzo},
  title   = {rizzo-pii: local reversible PII anonymization for Italian legal text},
  year    = {2026},
  url     = {https://huggingface.co/rizzoaiacademy/rizzo-pii-0.3B},
  note    = {mmBERT/ModernBERT token classification, 22 PII categories}
}