sandsiv-278-pii-ch

sandsiv-278-pii-ch

Multilingual PII named-entity recognition model for Swiss-domain texts, fine-tuned from microsoft/mdeberta-v3-base (278M parameters, fp32).

Detects 19 PII entity types in English, Italian, German and French, with a focus on Swiss-specific identifiers:

AHVNUMBER (Swiss social security), IBAN, CREDITCARD, PHONE, EMAIL, FULLNAME, ADDRESS, ZIPCODE, CITY, CANTON, DATEOFBIRTH, DATE, TIME, AGE, GENDER, ORG, DOCID, PLATE, LANDREGISTRY

Used in production

This model powers the PII detection engine of anonymo.io — a local-first desktop app for document sanitization built for Swiss FADP/GDPR compliance. All inference runs on-device; no document content ever leaves the machine. Built by SandSIV.

Intended use

PII detection as a building block for anonymization pipelines on Swiss texts (e.g. customer feedback, support tickets, administrative documents).

Usage

from transformers import pipeline

ner = pipeline(
    "token-classification",
    model="sandsiv/sandsiv-278-pii-ch",
    aggregation_strategy="simple",
)
ner("Il numero AVS di Anna Bianchi è 756.1234.5678.97.")
# -> FULLNAME 'Anna Bianchi', AHVNUMBER '756.1234.5678.97'

Note: the model is stored in fp32. Do not load it with dtype="auto" from code paths that would cast to fp16 before inference on CPU/MPS.

ONNX export (production deployment)

The onnx/ folder contains the exported model, verified to match the PyTorch reference (see evaluation below):

Inference requires only onnxruntime + tokenizers (no PyTorch/transformers). Session inputs: input_ids, attention_mask. Outputs: per-token logits (39 BIO labels — see config.json id2label).

import numpy as np
import onnxruntime as ort
from tokenizers import Tokenizer

tok = Tokenizer.from_file("tokenizer.json")  # from onnx/ or the repo root
tok.enable_truncation(128)
tok.enable_padding()

sess = ort.InferenceSession("onnx/model-int8.onnx")
enc = tok.encode_batch(["Il numero AVS di Anna Bianchi è 756.1234.5678.97."],
                       is_pretokenized=False)
logits = sess.run(["logits"], {
    "input_ids": np.array([e.ids for e in enc], dtype=np.int64),
    "attention_mask": np.array([e.attention_mask for e in enc], dtype=np.int64),
})[0]

Post-processing hint: merge adjacent same-label spans and trim boundary punctuation for clean entity strings.

Training data

100% synthetic data (generator ch-pii-synth-0.3.0 + targeted supplement), no real personal data was used. 29'000 examples, balanced across the 4 languages (7'250 each), 232+ templates, split 90/5/5 stratified by language.

Evaluation

Cross-generator robustness (8'000 samples from ai4privacy/pii-masking-200k, binary char-level PII detection): P 0.905 / R 0.552 / F1 0.685 overall — recall 0.891 on entity types covered by training (EMAIL, names, dates, addresses, phones, companies), 0.32 on types never trained (user-agent strings, crypto addresses, IPs, job titles). Precision is a lower bound: ai4privacy masks only a subset of PII per text.

Limitations

• Trained on synthetic data only: expect a quality gap on real-world text whose style differs from the generator's templates.
• Known weakness: document IDs with a letter prefix (e.g. S1234567A) may be split or partially mislabeled as PLATE.
• Entities spanning multiple SentencePiece subwords can occasionally be returned as adjacent fragments when using aggregation_strategy="simple"; merging adjacent same-label spans fixes this.
• AGE is the weakest entity type (test F1 ≈ 0.95).

Training configuration

lr 2e-5 · batch 16 · 4 epochs · weight decay 0.01 · warmup 10% · max length 128 · best checkpoint on validation entity-level F1 · Apple Silicon (MPS)

License

MIT (same as the base model).