maschera-ch-v63b

Use it: MASCHERA is the free desktop app (Windows, macOS, Linux, Docker) that runs this model to mask and anonymise personal data in Swiss documents before they go to an AI service. Source: github.com/semprerudi/maschera.

Token classification for personal data in Swiss documents in the three national languages German, French and Italian, as well as in English. Rumantsch is not supported.

It labels spans that carry personal data so they can be replaced by placeholders before a text is handed to an AI service.

Sehr geehrte Frau Brülhart      ->   Sehr geehrte Frau [FULLNAME_1]
AHV 756.1234.5678.97            ->   AHV [AHVN13_1]
Lindenstrasse 12, 3011 Bern     ->   [STREET_1] [BUILDINGNUM_1], [ZIPCODE_1] [CITY_1]

Where it is used

This model is the recognition stage of MASCHERA, a local masking tool for Swiss documents. MASCHERA fetches or bundles exactly this model; it runs on your own device and sends nothing anywhere.

• Website: www.maschera.ch
• Source code and downloads: github.com/semprerudi/maschera

What it is

Base model	jhu-clsp/mmBERT-base (ModernBERT), 22 layers, 308M parameters
Task	token classification, BIO
Labels	91 BIO labels over 45 tags, frozen contract b7a96dc9
Context	512 tokens; longer documents are windowed by the tool
Weights	model.safetensors, 1.2 GB
Training data	synthetic only — it has never seen a real document

The tag set covers Swiss identifiers with checksums (AHV, UID, IBAN, QR reference, GLN), addresses, names, dates, and administrative references. The full taxonomy lives in the repository.

How good it is

Two measurements. Both are here, not only the friendlier one.

	synthetic	gold documents
Size	2000 documents, 29,210 gold spans	9 documents, 280 gold spans
Leak rate	0.02 % (13 leaks)	1.27 % (6 leaks)
Micro-F1	0.997	0.736
Over-masking	84	95
Reproducible	yes	no

The synthetic figure is the reproducible one. Measured with MASCHERA 1.0.1 (python3 tools/mess_synthetisch.py --model runs/ch-v63b --n 2000); the documents depend on the templates and the generator, so the numbers change if those change. The documents are generated from the templates and nomenclatures in the repository; the values come from a holdout zone, so no value in the measurement was seen during training. That separates "recognises names" from "memorised a list" — but it does not replace real documents.

⚠️ The gold documents cannot be recomputed by anyone. They contain real personal data and are not published. There are nine, hand-marked. It is the most honest figure available and the least comfortable one.

⚠️ This test set resolves very little. Earlier evaluations on an older version of the test set (four training runs, each evaluated after epoch 2 and after epoch 3, eight evaluations in total) ranged from 0.769 to 0.833 micro-F1. Two models less than about 0.06 micro-F1 apart (roughly ±0.03) cannot be told apart by it. The 0.736 above was measured on the current test set and is not directly comparable to that range.

⚠️ Over-masking is not a leak. One word too many costs readability; one word too few costs personal data. The model is tuned towards the second.

What this means in practice

The leak rate on real documents is around one percent, not zero. Review the masked text before you pass it on. This is pseudonymisation, not anonymisation: as long as the dictionary exists the step is reversible, and under Swiss data protection law the data remain personal data.

Intended use

Masking personal data in Swiss administrative and business correspondence before it is given to a third-party AI service, with a human reviewing the result. It is a preparation step, not a compliance guarantee.

Not intended for unattended redaction, automated decisions, or as the sole safeguard before publishing a document.

How it was trained

Trained on synthetic data generated from the templates and nomenclatures of the repository. Values are drawn from open Swiss administrative sources (see Data sources below); no real document was used.

Epochs	3
Learning rate	3e-5, linear, warmup 0.1
Batch	8 × 2 (gradient accumulation)
Weight decay	0.01
Precision	bf16
max_length	512
Training seed	4711

Read from the run's training_args.bin, not copied from defaults.

⚠️ The seed of the training set is not recorded anywhere. The generator writes it into the header of the dataset it produces, but that file is not published. The path is reproducible; the exact run is not quite.

The generator, the templates and the nomenclatures are in the repository, so a comparable model can be built from scratch.

Data sources

The nomenclatures come from open Swiss government data. All of them require attribution:

• Surnames, given names — Federal Statistical Office (BFS)
• Localities with postal codes, streets — Federal Office of Topography (swisstopo)
• Places of origin — Federal Office of Justice, eCH-0135
• Company names — Zefix / EHRA via LINDAS, terms: Provide-the-Source

Limitations

• Trained on synthetic documents. Real correspondence is messier.
• The 512-token window is handled by the tool, not by the model. Used directly, long documents must be windowed by the caller.
• Recognition is statistical and incomplete; it misses things and masks too much when in doubt.
• Swiss-specific. The tag set and the anchors assume Swiss formats and the three official languages plus English. Rumantsch is not supported — neither the patterns nor the training data cover it.

Licence

Model weights: MIT.

Built on the idea of rizzo-pii by Simone Rizzo (MIT). No affiliation with that project.