Morpheus name model 6.0 — the name detector of the Morpheus de-identification engine for Greek clinical text (a GLiNER model fine-tuned on 7,500 synthetic clinical notes)

The Morpheus name model is a multilingual GLiNER token-classification model fine-tuned to find patient-identifying information in Greek clinical free text. It is the name-detection layer of Morpheus 6.0, the de-identification engine of ClinicalMiner (frozen 18 October 2026 as ClinicalMiner 1.7.0, pipeline fingerprint 9c833fe9e9fd4a7c).

Base model: urchade/gliner_multi-v2.1 (Apache-2.0), revision 443d26d654e0324125a96bebd8e796c14ff2efe6.

Training data: 7,500 synthetic clinical notes written for this purpose with their answer keys — 6,000 in Greek across the specialties of a general hospital (discharge summaries, operative notes, clinic letters, laboratory and pathology reports, emergency-department notes, in ordinary and all-capitals typing) and 1,500 in English — produced by the first author with a large language model under written style instructions, each note carrying its own span annotations for eight kinds of identifier: patient names, staff names, dates, hospital and record numbers, telephone numbers, addresses, hospital names and contact details. No real patient data was used at any stage. The notes were split into 383,032 training segments from 7,125 notes, with 375 notes held back.

Training: three passes with the GLiNER trainer: learning rate 5e-6 for the encoder and 1e-5 for the other parameters, weight decay 0.01, linear schedule with 10% warm-up, focal loss (alpha 0.75, gamma 2), batch size 8, seed 7; 143,500 optimisation steps in 262 minutes on one NVIDIA RTX 3090. Held-back loss 0.1755 after the first pass, 0.1414 after the second, 0.1423 after the third; training loss 3.10 to 0.047.

What is in the record: the exported ONNX graph of the fine-tuned model with its tokenizer and configuration, as the engine runs it (CPU, fp32). The export was proved equivalent to the training-time PyTorch weights on 10,372 Greek lines (0 decision differences, maximum score drift 4e-6) and 643 whole documents (0 different). The PyTorch checkpoint is not included.

Measured behaviour: inside the engine, against the previous public model, the trained model finds the same names and destroys the same or fewer clinical words on thirteen synthetic Greek development corpora. Alone, asked for persons and locations, it caught 1,269 of 1,269 names on a 500-note benchmark corpus and 748 of 748 on a sealed 200-note corpus, altering no protected clinical entry. The figures are reported in the authors' papers.

Related: the benchmark corpus of 500 synthetic Greek discharge summaries, doi:10.5281/zenodo.22824894 (CC BY 4.0); the benchmark paper "Morpheus: a fully local, defense-in-depth de-identification engine for Greek clinical text — a benchmark evaluation" (arXiv, October 2026; under review). The training notes live in the authors' project repository and are available on request together with the weights.

This model runs inside the Morpheus application for macOS, available at https://github.com/prorris/Morpheus.