choosing-openmed-models

Choosing OpenMed Models

OpenMed ships a registry of clinical and biomedical NER models grouped into 12 categories. Never hardcode a model list — query the registry at runtime so your code stays correct as models are added. This skill helps you go from "I need to find diseases in Spanish discharge notes" to a concrete model key.

When to use

• The user knows the task (find diseases / tumors / PHI) but not the model.
• You need the right PII model for a language (es, fr, de, …).
• You want to filter models by size, task, or tier before loading.
• You want to inspect a model's labels, params, and license first.

Once you have a key, hand off to loading-openmed-models to load it.

Install

pip install openmed         # registry queries work without the [hf] extra

Quick start: browse categories, then pick

import openmed

# 1) The 12 categories
openmed.list_model_categories()
# ['Medical', 'Privacy', 'Anatomy', 'Hematology', 'Chemical', 'Disease',
#  'Genomics', 'Oncology', 'Species', 'Pathology', 'Pharmaceutical', 'Protein']

# 2) Models in a category -> list[ModelInfo]
for m in openmed.get_models_by_category("Disease"):
    print(m.model_id, "|", m.size_category, "|", m.entity_types)

# 3) Inspect one model before loading
info = openmed.get_model_info("OpenMed/OpenMed-NER-DiseaseDetect-BigMed-278M")
print(info.display_name, info.task, info.param_count, info.license)

get_models_by_category and get_all_models return ModelInfo objects. get_all_models() returns a dict[str, ModelInfo] keyed by registry key.

What ModelInfo tells you

Every model exposes (real attributes):

model_id          # HF repo id, e.g. "OpenMed/OpenMed-NER-DiseaseDetect-BigMed-278M"
display_name      # human-friendly name
category          # one of the 12 categories
specialization    # e.g. "disease entity detection"
entity_types      # list[str] of labels the model emits, e.g. ["DISEASE", ...]
size_category     # "Tiny" | "Small" | "Medium" | "Large" | "XLarge"
recommended_confidence   # suggested confidence_threshold for this model
family            # "NER" | "PII" | ...
task              # "token-classification"
languages         # e.g. ["en"], ["es"]
param_count       # e.g. 278000000
license           # e.g. "apache-2.0"