leonx1995/docudis-ner-xlmr

Docudis NER (xlmr_ner_docudis)

The name-recognition model of Docudis, which replaces personal details in documents with labels on the user's device. It finds people (PER), organisations (ORG), places and addresses (LOC) and dates (DATE) in English, French and Spanish text.

• Format: ONNX, int8-quantized (model_quantized.onnx), with its SentencePiece tokenizer (tokenizer.json).
• Inputs: windows of up to 256 tokens; Docudis runs overlapping windows with a stride of 32 and keeps spans whose mean probability is at least 0.5. model.json describes the labels and settings.
• Code: inference runs through docudis-ner (Apache-2.0).

How it was made

This model is a full fine-tune of Davlan/xlm-roberta-base-ner-hrl, which is itself fine-tuned from FacebookAI/xlm-roberta-base. The label set and order are unchanged.

The fine-tuning data, in English, French and Spanish in equal shares:

Share	Material
60 %	Document templates written for this purpose, filled with generated names, addresses and identifiers
25 %	Official company announcements: BODACC (France), BORME (Spain), The Gazette (United Kingdom)
15 %	Complete documents written for this purpose, with hand-marked entities

The pipeline is in docudis-ner/training. The training text is not distributed with the model.

Data from official gazettes

• BODACC: Direction de l'information légale et administrative (DILA), under the Licence Ouverte 2.0.
• BORME: Agencia Estatal Boletín Oficial del Estado (www.boe.es).
• The Gazette (www.thegazette.co.uk): contains public sector information licensed under the Open Government Licence v3.0, Crown copyright.

These announcements name company officers and insolvency practitioners, as published by law in those gazettes.

The base model's training data

The model card of Davlan/xlm-roberta-base-ner-hrl lists its training corpora: CoNLL-2003 (English, German), CoNLL-2002 (Spanish, Dutch), Europeana Newspapers (French), ANERcorp (Arabic), MSRA (Chinese), I-CAB (Italian), the Latvian NER dataset and Paramopama + Second HAREM (Portuguese). Some of these corpora are distributed under research-only terms. This model inherits the base model's weights, so evaluate those terms for your use.

Limitations

No automatic detection finds every personal detail, and this model is no exception. It is tuned for business and personal documents in English, French and Spanish; other languages fall back to what the base model learned. Always review its output before sharing a document.

License

Copyright 2026 Pengda Xia (stonetech). Licensed under the Academic Free License 3.0.

This model is a Derivative Work of Davlan/xlm-roberta-base-ner-hrl (Academic Free License 3.0), modified by stonetech by fine-tuning it on the data above and quantizing it. XLM-RoBERTa is licensed under the MIT License, Copyright (c) Facebook, Inc. and its affiliates.