Loading OpenMed Models

OpenMed models download once from the Hugging Face Hub into a local cache, then run fully on-device — no network, no telemetry. This skill covers how to load a model, reuse it across many calls without reloading weights, point at a local copy, and run offline.

When to use

• You are about to run NER repeatedly and want to load the model once.
• You need to control where weights are cached (cache_dir) or force CPU/GPU.
• You must run offline in a locked-down or air-gapped environment.
• You are choosing between a registry key, a full HF id, or a local directory.

For which model to load, see choosing-openmed-models. To actually run it, see extracting-clinical-entities.

Install

pip install "openmed[hf]"   # adds Hugging Face transformers + hub download

The three ways to name a model

analyze_text, extract_pii, load_model, and ModelLoader.load_model all accept the same model_name in three forms:

Form	Example	Notes
Registry key	"disease_detection_superclinical"	Short, resolved via the bundled registry.
Full HF id	"OpenMed/OpenMed-NER-DiseaseDetect-BigMed-278M"	Anything org/name; downloaded from the Hub.
Local path	"/models/my-openmed-ner"	An existing directory; loaded with local_files_only=True.

A bare name without / is prefixed with the default org (OpenMed). An existing local path is detected automatically and never hits the network.

Quick start: load and reuse a loader

The single most important pattern — build one ModelLoader, pass it everywhere. The loader caches models, tokenizers, and pipelines in memory, so the second call is instant.

import openmed
from openmed import ModelLoader, OpenMedConfig

# One loader, reused across calls. Weights load on the first call only.
loader = ModelLoader()

notes = [
    "Patient prescribed 500 mg metformin for type 2 diabetes.",
    "History of myocardial infarction; started on atorvastatin.",
]

for note in notes:
    result = openmed.analyze_text(
        note,
        model_name="disease_detection_superclinical",
        loader=loader,          # <-- reuse; no reload on subsequent calls
        output_format="dict",
    )
    print(result.entities)

Without loader=, each analyze_text call constructs a fresh ModelLoader. The underlying Hugging Face cache still prevents re-downloads, but you pay to re-instantiate the pipeline — avoid that in loops and services.

Load weights directly

When you want the raw model/tokenizer (e.g. to inspect config or build a custom pipeline):

from openmed import load_model

bundle = load_model("disease_detection_superclinical")
model     = bundle["model"]
tokenizer = bundle["tokenizer"]
config    = bundle["config"]

load_model(model_name, config=None, **kwargs) is a thin convenience wrapper that builds a ModelLoader and calls loader.load_model(...). For reuse, prefer constructing the loader yourself:

loader = ModelLoader()
bundle = loader.load_model("disease_detection_superclinical")
# Second call returns the cached bundle (no reload):
bundle2 = loader.load_model("disease_detection_superclinical")
# Force a fresh load if you replaced files on disk:
fresh = loader.load_model("disease_detection_superclinical", force_reload=True)

Configure the cache, device, and org

OpenMedConfig is a dataclass. Pass it to ModelLoader(config=...).

from openmed import ModelLoader, OpenMedConfig

config = OpenMedConfig(
    cache_dir="/data/openmed-cache",   # default: ~/.cache/openmed
    device="cpu",                       # None = auto-detect
    default_org="OpenMed",              # prepended to bare model names
    hf_token=None,                      # or set env HF_TOKEN for private repos
)
loader = ModelLoader(config)

Relevant OpenMedConfig fields: cache_dir, device, default_org, hf_token, timeout (default 300s), backend (None auto / "hf" / "mlx"), log_level. hf_token falls back to the HF_TOKEN environment variable.

First-run download, then fully offline

1. First run (online): the model is fetched from the Hub into cache_dir.
2. Every run after: transformers serves from cache with no network call.

To guarantee no network access (air-gapped, CI, PHI environments), set the standard Hugging Face offline switch before importing:

export HF_HUB_OFFLINE=1
export TRANSFORMERS_OFFLINE=1

Or vendor the model and pass a local path — that path is loaded with local_files_only=True and never contacts the Hub:

result = openmed.analyze_text(note, model_name="/models/openmed-disease-ner")

To pre-warm a cache for offline use, run one inference (or load_model) once with network access, then disable it.

Check a model's maximum sequence length

Useful before chunking long documents:

from openmed import get_model_max_length, ModelLoader

loader = ModelLoader()
max_len = get_model_max_length("disease_detection_superclinical", loader=loader)
print(max_len)   # e.g. 512 — None if it can't be inferred

get_model_max_length(model_name, *, config=None, loader=None) delegates to loader.get_max_sequence_length(model_name). Pass the same loader you use for inference so the tokenizer is loaded only once.

Free memory when done

The loader holds models in RAM until released:

loader.unload_model("disease_detection_superclinical")  # drop one model
loader.unload_all_models()                               # drop everything
loader.loaded_models()                                   # inspect what's cached

Hand-off to / from OpenMed

• From choosing-openmed-models: that skill yields a model key or HF id; feed it straight into ModelLoader.load_model(...) or as model_name=.
• To extracting-clinical-entities: pass your reused loader= into openmed.analyze_text(...) so a long batch loads weights exactly once.
• To de-identification: openmed.extract_pii(..., loader=loader) and openmed.deidentify(..., loader=loader) accept the same loader — share one loader across NER and PHI steps in a pipeline.

loader = ModelLoader(OpenMedConfig(cache_dir="/data/openmed-cache"))
phi   = openmed.deidentify(note, method="mask", loader=loader)
ner   = openmed.analyze_text(phi.deidentified_text, loader=loader)

Edge cases & gotchas

• pip install openmed alone is not enough to download models — add the [hf] extra (or have transformers + huggingface_hub installed). ModelLoader raises ImportError with an install hint if transformers is missing.
• Local path vs registry key collision: if a bare name happens to exist as a directory, the local path wins. Use an absolute path to be explicit.
• force_reload=True is required after you overwrite files in a local model directory; otherwise the in-memory cache is served.
• Private repos need hf_token (or HF_TOKEN) and HF_HUB_OFFLINE unset for the first download.
• No PHI in the cache path or logs. Cache model weights, never patient text. cache_dir should not live inside a PHI data directory.
• Permissive licensing only. OpenMed models are Apache-2.0. Do not stage UMLS/SNOMED/CPT/MIMIC/i2b2/n2c2 assets in the cache — those stay out-of-process under the user's own license.

Standards & references

• Hugging Face Hub caching & offline mode: https://huggingface.co/docs/huggingface_hub/guides/manage-cache and https://huggingface.co/docs/transformers/installation#offline-mode
• OpenMed model org on the Hub: https://huggingface.co/OpenMed