--- license: apache-2.0 language: - es - en - fr - de - it - nl - sv library_name: openmed pipeline_tag: token-classification base_model: fastino/gliner2.5-multi-v1 datasets: - bigbio/meddocan - gretelai/synthetic_pii_finance_multilingual - elenanereiss/german-ler tags: - medical - clinical - pii - de-identification - anonymization - privacy - multilingual - gliner2 - gliner - openmed model-index: - name: GLiNER2.5-PII-Multilingual-v1 results: - task: type: token-classification name: PII masking (character spans) dataset: name: MEDDOCAN (91 held-out Spanish clinical case reports) type: bigbio/meddocan split: held-out metrics: - type: f2 value: 98.57 name: Character F2 (threshold 0.15) - type: f1 value: 97.60 name: Exact-span F1 (threshold 0.15) - type: recall value: 98.94 name: Critical identifiers fully masked (threshold 0.15) - task: type: token-classification name: PII masking (character spans) dataset: name: Gretel synthetic PII finance multilingual (588 held-out documents, 7 languages) type: gretelai/synthetic_pii_finance_multilingual split: held-out metrics: - type: f2 value: 76.95 name: Character F2, macro over 7 languages (threshold 0.15) - task: type: token-classification name: PII masking (character spans) dataset: name: German LER (107 person names) type: elenanereiss/german-ler split: train metrics: - type: recall value: 61.68 name: Names fully masked (threshold 0.15) ---

OpenMed

OpenMed GLiNER2.5 PII Multilingual v1

OpenMed on Hugging Face OpenMed SDK on GitHub pip install openmed License Apache-2.0

**OpenMed's open model for masking personal information in clinical text, in several languages.** It finds the identifiers in Spanish clinical notes (names, record and license numbers, dates, ages, addresses, phone numbers, e-mails) and returns each one as an exact character span with a score, ready to mask. It also covers English, French, German, Italian, Dutch and Swedish, tested on synthetic finance documents. Trained and evaluated by [OpenMed](https://huggingface.co/OpenMed). Apache-2.0, 287.4M parameters, runs on CPU. | Developed by | [OpenMed](https://huggingface.co/OpenMed) | |---|---| | Task | PII masking, one generic label: `personally identifying information` (the prompt it was trained with) | | Output | `text`, `start`, `end`, `label`, `score`; `text[start:end]` is the exact span | | Parameters | 287.4M | | Default threshold | 0.15 | | Languages | Spanish (clinical text); English, French, German, Italian, Dutch, Swedish (tested on synthetic finance text) | | License | Apache-2.0 | ## About OpenMed OpenMed builds open models and open-source tools for clinical and biomedical text that run on your own hardware. - **[2,300+ open models on Hugging Face](https://huggingface.co/OpenMed)** for clinical NER, de-identification and more. - **[The OpenMed Python SDK](https://github.com/maziyarpanahi/openmed)**, Apache-2.0: `pip install openmed` for de-identification, clinical NER, and FHIR and HL7 tooling, all local. - New models every week: follow [OpenMed on Hugging Face](https://huggingface.co/OpenMed) and [@OpenMed_AI on X](https://x.com/OpenMed_AI). ## Quick start ```bash pip install -U "openmed[gliner2]" # OpenMed 3.2 or later ``` ```python from openmed import OpenMedConfig, analyze_text text = ( "Paciente: Lucía Martín Herrera, 58 años, NHC 2847193. " "Ingresa el 14/03/2025 en el Hospital Universitario Río " "Tajo por disnea de esfuerzo." ) result = analyze_text( text, model_id="OpenMed/GLiNER2.5-PII-Multilingual-v1", labels=["personally identifying information"], threshold=0.1, confidence_threshold=0.15, max_length=120, config=OpenMedConfig(use_medical_tokenizer=False), ) for e in result.entities: print(e.start, e.end, e.text, f"{e.confidence:.2f}") masked = text for e in sorted(result.entities, key=lambda e: e.start, reverse=True): masked = masked[: e.start] + "[PII]" + masked[e.end :] print(masked) ``` ```text 10 30 Lucía Martín Herrera 1.00 32 39 58 años 0.99 45 52 2847193 0.99 65 75 14/03/2025 0.98 82 113 Hospital Universitario Río Tajo 0.99 Paciente: [PII], [PII], NHC [PII]. Ingresa el [PII] en el [PII] por disnea de esfuerzo. ``` ### The settings - `labels=["personally identifying information"]` is the prompt the model was trained with. Keep it as written. Every span comes back with this label: the model marks what to mask, not which kind of identifier it is. - `threshold=0.1` lets the model propose spans down to 0.1, and `confidence_threshold=0.15` keeps those scoring 0.15 or more. This is the decoding behind the evaluation below; the threshold was chosen on the tuning data to favor recall. Raise `confidence_threshold` for fewer, surer spans, at the cost of leaving more identifiers unmasked. - `max_length=120` makes `analyze_text` group sentences into pieces of at most 480 characters, close to the windows used in training and evaluation (at most 438 UTF-8 bytes). It only matters for long texts. - `use_medical_tokenizer=False` keeps the model's own span boundaries. With the SDK default, spans are remapped onto its own word tokens: in example 3 it returns *46002* and *Valencia* as one span, *46002 Valencia*; in example 4 it widens *Carmen* to *hija Carmen*. ### Long notes, many notes, GPU - Pass a whole note: `analyze_text` splits it into sentences, groups them into pieces and returns offsets into the original text. - The model loads once and stays cached, so call `analyze_text` in a loop for many notes. - On a GPU, pass `config=OpenMedConfig(device="cuda", use_medical_tokenizer=False)`. ## Evaluation **Spanish clinical case reports (main result):** 91 case reports from the MEDDOCAN corpus, from 90 article groups held out of training and tuning, with 2,065 annotated identifiers. 946 of them are critical: names, ID numbers, e-mails, phone and fax numbers, relatives and street addresses. The case text is authentic; the identifiers in it were substituted or added by the corpus authors. All cases come from the MEDDOCAN authors' training file (the official test set was not used), so these numbers are not comparable to published MEDDOCAN results. Scores are counted per character in the original text: F2 weighs recall twice as much as precision, because a missed identifier costs more than an extra mask. An identifier is fully masked when every one of its characters is. Three more views, on the same frozen model: **588 synthetic finance documents** in seven languages (English, French, German, Italian, Spanish, Dutch and Swedish, 84 each) from 12 document types held out of training and tuning, where dates, times and company names are outside the mask scope (macro F2 is the mean of the seven languages); **107 person names in German court decisions**, counted as fully masked or not (the gold marks names only, so there is no precision); and **24 short clinical notes with no personal data** (12 Spanish, 6 German, 6 French), written as controls, counted as masked if anything in them is masked. | Model | Spanish character F2 | Spanish exact-span F1 | Spanish critical identifiers fully masked | Finance macro F2 (7 languages) | German names fully masked | Clean notes masked (of 24) | |---|---:|---:|---:|---:|---:|---:| | **OpenMed GLiNER2.5 PII Multilingual v1** (this model), threshold 0.15 | **98.57** | **97.60** | **98.94** | **76.95** | **61.68** | **0** | | Earlier OpenMed PII model (EmbeddingGemma2 PII Multilingual 270M v2, frozen reference) | 84.64 | 43.12 | 81.71 | 65.27 | 63.55 | 1 | | Before OpenMed fine-tuning: [fastino/gliner2.5-multi-v1](https://huggingface.co/fastino/gliner2.5-multi-v1), same prompt, threshold 0.5 | 35.77 | 37.77 | 51.27 | 38.52 | 53.27 | 0 | OpenMed's fine-tuning raised Spanish character F2 from 35.77 to **98.57**. Against the earlier trained reference it gained **+13.92 points** (95% CI +12.80 to +15.10, paired bootstrap over the 90 article groups, 2,000 resamples), and the share of critical identifiers fully masked rose by +17.23 points (95% CI +14.74 to +19.69). The reference was chosen on the tuning data from three earlier OpenMed PII models, the threshold (0.15) was chosen on the tuning data too, and these cases were scored once, with the model, threshold and scoring code frozen first. The model was trained on 303 other MEDDOCAN case reports and 2,730 synthetic finance documents. **What it still misses:** 19 of the 91 Spanish case reports kept at least one identifier character unmasked, and 9 kept part of a critical identifier (the reference: 91 and 82). It masked 0.08% of the characters that are not identifiers (the reference: 0.27%). ### By identifier type (Spanish case reports) | Identifier (MEDDOCAN type) | Critical | Annotated | Fully masked | Share fully masked | |---|:---:|---:|---:|---:| | place (city, province, postcode) (`TERRITORIO`) | | 332 | 328 | 98.80 | | date (`FECHAS`) | | 229 | 229 | 100.00 | | patient age (`EDAD_SUJETO_ASISTENCIA`) | | 189 | 187 | 98.94 | | patient name (`NOMBRE_SUJETO_ASISTENCIA`) | yes | 185 | 185 | 100.00 | | clinician name (`NOMBRE_PERSONAL_SANITARIO`) | yes | 182 | 182 | 100.00 | | patient sex (`SEXO_SUJETO_ASISTENCIA`) | | 172 | 172 | 100.00 | | street address (`CALLE`) | yes | 156 | 155 | 99.36 | | country (`PAIS`) | | 131 | 131 | 100.00 | | patient ID or record number (`ID_SUJETO_ASISTENCIA`) | yes | 105 | 101 | 96.19 | | e-mail (`CORREO_ELECTRONICO`) | yes | 85 | 84 | 98.82 | | clinician license number (`ID_TITULACION_PERSONAL_SANITARIO`) | yes | 84 | 84 | 100.00 | | insurance number (`ID_ASEGURAMIENTO`) | yes | 71 | 71 | 100.00 | | relatives (`FAMILIARES_SUJETO_ASISTENCIA`) | yes | 44 | 40 | 90.91 | | hospital (`HOSPITAL`) | | 42 | 41 | 97.62 | | encounter ID (`ID_CONTACTO_ASISTENCIAL`) | yes | 18 | 18 | 100.00 | | institution (`INSTITUCION`) | | 17 | 13 | 76.47 | | phone number (`NUMERO_TELEFONO`) | yes | 11 | 11 | 100.00 | | profession (`PROFESION`) | | 6 | 1 | 16.67 | | fax number (`NUMERO_FAX`) | yes | 5 | 5 | 100.00 | | other patient detail (`OTROS_SUJETO_ASISTENCIA`) | | 1 | 0 | 0.00 | The weakest types (at least five annotations, under 95% fully masked): profession, 1 of 6; institution, 13 of 17; relatives, 40 of 44. ### Synthetic finance documents by language | Character F2 | English | French | German | Italian | Spanish | Dutch | Swedish | |---|---:|---:|---:|---:|---:|---:|---:| | **OpenMed GLiNER2.5 PII Multilingual v1** | 91.46 | 78.94 | 80.90 | 82.11 | 75.16 | 60.66 | 69.40 | | Earlier trained reference | 77.28 | 72.02 | 65.25 | 70.42 | 65.16 | 47.34 | 59.41 | | Before fine-tuning | 42.39 | 34.80 | 40.77 | 43.40 | 42.78 | 29.34 | 36.20 | Dutch is its weakest finance language (60.66). Across the seven languages names are the most common finance identifier, and 764 of the 1,404 names (54.42%) were fully masked. These documents are synthetic, so the numbers show transfer across languages, not performance on real records. **German names are the one view where it does not beat the reference:** 61.68% of the 107 names fully masked against 63.55% for the earlier reference (77.47% and 81.85% of name characters). ## The OpenMed multilingual PII models Same training data, same frozen comparison, same views. GLiNER2.5 is ahead on every scored view, and neither masks any of the 24 clean notes; mmBERT runs with plain transformers and no extra library. | Model | Parameters | Default threshold | Quick start | Spanish character F2 | Spanish critical identifiers fully masked | Finance macro F2 (7 languages) | German names fully masked | Clean notes masked (of 24) | |---|---:|---:|---|---:|---:|---:|---:|---:| | [OpenMed GLiNER2.5 PII Multilingual v1](https://huggingface.co/OpenMed/GLiNER2.5-PII-Multilingual-v1) (this model) | 287.4M | 0.15 | OpenMed SDK | **98.57** | **98.94** | **76.95** | **61.68** | **0** | | [OpenMed mmBERT PII Multilingual Base v1](https://huggingface.co/OpenMed/mmBERT-PII-Multilingual-Base-v1) | 307.5M | 0.35 | transformers | 98.12 | 97.89 | 74.04 | 36.45 | 0 | | Earlier OpenMed PII model (EmbeddingGemma2 270M v2, reference, not released) | | | | 84.64 | 81.71 | 65.27 | 63.55 | 1 | ## Examples Real outputs of this model through the OpenMed SDK, with the default threshold 0.15, on CPU. The texts are synthetic, written for this card: every name, number, address, e-mail and hospital is invented. Spans to mask are in bold with their score. 1. *Admission note (Spanish):* Paciente: **Lucía Martín Herrera**1.00, **58 años**0.99, NHC **2847193**0.99. Ingresa el **14/03/2025**0.98 en el **Hospital Universitario Río Tajo**0.99 por disnea de esfuerzo. 2. *Signing physician (Spanish):* Informe firmado por la Dra. **Elena Sanz Prieto**1.00, número de colegiado **284512706**1.00, Servicio de Neumología. Correo: **esanz.neumo@hospitalriotajo.es**1.00. 3. *Address and phone (Spanish):* Vive con su marido en la calle **Pintor Sorolla 23, 4º B**0.85, **46002**0.16 **Valencia**0.16. Teléfono de contacto: **961 234 567**1.00. 4. *Relative and profession (Spanish):* **Varón**1.00 de **71 años**0.99, NASS **28 41235678 09**0.99, que acude acompañado de su hija **Carmen**0.98, enfermera de profesión. 5. *English discharge (English):* Discharge summary for **John Patrick Miller**1.00, DOB **04/17/1962**1.00, MRN **00482913**0.96. Follow-up with **Dr. Sarah Chen**0.93, phone **(617) 555-0142**0.96. 6. *French referral (French):* **Mme Claire Dubois**1.00, née le **3 mars 1958**1.00, domiciliée **12 rue des Lilas, 69003 Lyon**0.99, est adressée par le **Dr Antoine Mercier**0.89. 7. *German bank details (German):* **Herr Klaus Becker**0.98, geboren am **14.08.1971**0.99, wohnhaft in der **Hauptstraße 27, 10115 Berlin**0.83, IBAN **DE89 3704 0044 0532 0130 00**0.94. 8. *Italian tax code (Italian):* Il signor **Marco Bianchi**1.00, codice fiscale BNCMRC80A01F205X, residente in **Via Garibaldi 8, 20121 Milano**0.99, telefono **02 4567 8910**0.99. 9. *Dutch account holder (Dutch):* Rekeninghouder **Jan de Vries**0.99, **Keizersgracht 112**0.98, **1015 CW Amsterdam**0.57, e-mail **j.devries@voorbeeld.nl**1.00. 10. *No personal data (Spanish):* Se pauta amoxicilina 875 mg cada 8 horas durante 7 días y control de la tensión arterial en atención primaria. *(nothing masked)* **A longer Spanish note** (803 characters, 16 spans) in one `analyze_text` call: > **Mujer** de **46 años**, natural de **Toledo**, que acude a urgencias del **Hospital General Sierra Alta** el 2 de febrero de 2025 por fiebre y dolor lumbar. Antecedentes: hipertensión arterial en tratamiento con enalapril. Vive en la **avenida de la Constitución 51, 2º A**, **45004** **Toledo**, con su pareja, **Andrés Molina**. NHC **5530918**, NASS **45 10293847 22**. > Exploración: temperatura 38,7 °C, puñopercusión renal derecha positiva. Analítica con leucocitosis y proteína C reactiva elevada. Se diagnostica pielonefritis aguda y se inicia ceftriaxona intravenosa. > Evolución favorable, con alta a las 72 horas y control por su médica de familia, la Dra. **Beatriz Lozano Gil**, en el **centro de salud Los Almendros**. Remitida por: Dr. **Tomás Ibáñez Ruiz**, nº de colegiado **454503218**. Contacto: **tibanez@hgsierraalta.es**, teléfono **925 280 114**. ## Good to know - Across the six languages of examples 1 to 9, one call masks names, record and license numbers, dates, addresses, phone numbers, IBANs and e-mails. It also masks the patient's sex and age (*Varón*, *71 años* in example 4, *Mujer* in the longer note), which MEDDOCAN counts as identifiers. - Titles stay inside the span: *Dr. Sarah Chen*, *Dr Antoine Mercier*, *Herr Klaus Becker* (examples 5 to 7). - It leaves professions and relationship words: example 4 masks the daughter's name *Carmen* but not *hija* or *enfermera*, and example 3 leaves *marido*. Professions are its weakest type in the evaluation too (1 of 6 fully masked). - Some spans pass the 0.15 threshold narrowly: the postcode *46002* (0.16) and the city *Valencia* (0.16) in example 3. A higher threshold would leave them visible. - Example 10, a prescription with no personal data, comes back with nothing masked. - It misses the Italian tax code *BNCMRC80A01F205X* (example 8) and the written-out date *2 de febrero de 2025* in the longer note. ## Limitations - The main result is in-distribution: the 91 Spanish case reports are MEDDOCAN article groups held out of the same author training file the model learned from, and their identifiers were substituted or added by the corpus authors. It was not tested on the official MEDDOCAN test set or on real hospital notes. - Clinical training data is Spanish only. The other languages were measured on synthetic finance documents, not on clinical notes, and Portuguese and other languages were not trained or tested. - German names: 61.68% fully masked, below the earlier reference's 63.55%; in the synthetic finance documents 764 of the 1,404 names were fully masked. - One generic label. It does not say which kind of identifier a span is, so type-aware replacement (fake names for names, shifted dates for dates) needs another step. - Scores rank spans; they are not calibrated probabilities. - Long texts are cut into pieces, and an identifier cut across two pieces can be partly missed. - **It does not guarantee de-identification.** Even on the in-distribution Spanish test, 19 of 91 case reports kept part of an identifier. Have a person review the output before any text leaves your control, and follow the rules that apply to your data. - The base model's pretraining data is not public. - It finds text spans. It does not give medical advice and is not a medical device. ## Built on OpenMed fine-tuned this model from [fastino/gliner2.5-multi-v1](https://huggingface.co/fastino/gliner2.5-multi-v1) by Fastino (Apache-2.0), built with the open-source [GLiNER2](https://github.com/fastino-ai/GLiNER2) library. Training and tuning data: Spanish clinical case reports from the [MEDDOCAN corpus](https://zenodo.org/records/4279323) by Marimon et al. (2019, CC BY 4.0; Zenodo record 4279323), whose identifiers were substituted or added by the corpus authors, and the [Gretel synthetic PII finance multilingual dataset](https://huggingface.co/datasets/gretelai/synthetic_pii_finance_multilingual) (Apache-2.0). Evaluation only: person names from the [German LER corpus](https://huggingface.co/datasets/elenanereiss/german-ler) of German court decisions by Leitner et al. (CC BY 4.0). OpenMed's changes: every identifier type was mapped to one generic mask, and MEDDOCAN cases linked to earlier OpenMed Spanish data (27) were left out. Thank you to Fastino, to the MEDDOCAN authors, to Gretel and to the German LER authors for releasing their work openly. This model is released under Apache-2.0. No corpus text is included in this repository. ## Citation ```bibtex @misc{panahi2025openmedner, title={OpenMed NER: Open-Source, Domain-Adapted State-of-the-Art Transformers for Biomedical NER Across 12 Public Datasets}, author={Maziyar Panahi}, year={2025}, eprint={2508.01630}, archivePrefix={arXiv} } @inproceedings{marimon2019meddocan, title={Automatic De-identification of Medical Texts in Spanish: the {MEDDOCAN} Track, Corpus, Guidelines, Methods and Evaluation of Results}, author={Marimon, Montserrat and Gonzalez-Agirre, Aitor and Intxaurrondo, Ander and Rodriguez, Heidy and Lopez Martin, Jose Antonio and Villegas, Marta and Krallinger, Martin}, booktitle={Proceedings of the Iberian Languages Evaluation Forum (IberLEF 2019)}, series={CEUR Workshop Proceedings}, volume={2421}, year={2019} } @inproceedings{leitner2020german, title={A Dataset of German Legal Documents for Named Entity Recognition}, author={Leitner, Elena and Rehm, Georg and Moreno-Schneider, Julian}, booktitle={Proceedings of the 12th Language Resources and Evaluation Conference (LREC 2020)}, year={2020} } @misc{zaratiana2025gliner2, title={GLiNER2: An Efficient Multi-Task Information Extraction System with Schema-Driven Interface}, author={Urchade Zaratiana and Gil Pasternak and Oliver Boyd and George Hurn-Maloney and Ash Lewis}, year={2025}, eprint={2507.18546}, archivePrefix={arXiv} } ```

OpenMed · Hugging Face · GitHub · PyPI · X