anonimato-ner-ptbr (bertimbau-anm)

Named-entity model of Anonimato, a free desktop app that masks people, companies and addresses in Brazilian documents on the user's own computer before they are sent to an AI assistant. ONNX, int8 (per-channel dynamic quantization), 109 MB, runs on onnxruntime without torch.

Labels: PESSOA, ORG, LOCAL, ENDERECO (BIO). In the app the model is one part of a hybrid: addresses and CEP come mainly from rules, public bodies are filtered out of ORG, and every form of a name ("MARIA DOS SANTOS", "Sra. Santos") gets one marker.

Training

• Base: BERTimbau-base (MIT, Souza, Nogueira and Lotufo, 2020).
• Data: LeNER-Br train split (MIT, Luz de Araujo et al., 2018), plus 6,000 paragraphs of a synthetic corpus of petitions, contracts, powers of attorney, notices, e-mails, rulings and declarations with randomly generated people, addresses and companies. No real personal data.
• 2 epochs, lr 4e-5, batch 8, max length 128. Recipe: core/train/train_ner.py in the Anonimato repository.

Evaluation

Word-level recall through the whole Anonimato pipeline (rules + model), on a synthetic test corpus built from disjoint name lists and different phrasing: PESSOA 1.000, ENDERECO 1.000, EMPRESA 0.993, overall 0.999 (precision 0.968). On 24 hand-written fragments outside the generator: 0.956. Real documents were not measured yet; do not rely on the model alone for anonymisation in the sense of LGPD art. 12.

Licence

Apache-2.0. Attribution: BERTimbau (MIT) and LeNER-Br (MIT).