OpenJev healthcare router

A worked example you can run in three lines. It routes pharmacy messages: one intent, five multi-label topic flags and four safety gates, all answered in a single forward pass over the shared message.

It exists to show what 395 training examples and 38 seconds on one GPU buy you, and it is the model behind the numbers in the results.

from openjev import DecisionModel, Choice, Noul

model = DecisionModel.from_pretrained("s1lv3rj1nx/openjev-router-healthcare")

out = model.answer(
    "I need a refill on my thyroid tablets, and are you open on Sunday?",
    {
        "A_intent": Choice(
            instructions="Which single topic does this pharmacy message concern?",
            criteria={
                "refill": "asking us to dispense another fill of an existing prescription",
                "store_hours": "when a branch is open",
                "order_status": "where an existing order is",
            },
        ),
        "C_store_hours": Noul(instructions="Does this message ask about when a branch is open?"),
        "G_clinical": Noul(instructions="Does this describe a clinical symptom?"),
    },
)

out["A_intent"].label                         # 'refill'      p = 0.874
out["C_store_hours"].probabilities["true"]    # 1.000  — both topics caught
out["G_clinical"].probabilities["true"]       # 0.000

pip install git+https://github.com/S1LV3RJ1NX/openjev for the package. The checkpoint carries its own backbone name, prompt format and fitted temperatures, so there is nothing else to configure.

Measured against a commercial typed-decision API

Paired on the same 450 held-out test items, exact McNemar. Reproducible with scripts/compare_to_jev.py.

Per-question test accuracy, temperature-scaled, with bootstrap CIs:

Limitations, stated plainly

It is a demo on synthetic data. The task is synthetic pharmacy messages, not a deployment. Treat the numbers as evidence that the recipe works, not as a claim about your traffic.

G_pharmacy over-fires. False-positive rate 0.652. The slice is 427 positive to 23 negative, and 23 negatives is not enough to train a gate properly. It still beats the commercial API overall on that gate, which says more about the difficulty than about either model.

G_injection is the weakest gate, recall 0.583 against the API's 0.917, and it regressed from 0.750 under an earlier general checkpoint.

Do not use this for real clinical triage. It flags whether a message mentions clinical content so it can be escalated. It is not a medical device and makes no clinical judgement.

How it was made

hf download s1lv3rj1nx/openjev-encoder-general model.pt --local-dir checkpoints/general
uv run python scripts/train.py --task tasks/healthcare_router \
    --init-from checkpoints/general/model.pt --epochs 6 --bs 8

Starting from the general checkpoint instead of raw ModernBERT is worth +36 points for the same 38 seconds: intent 0.544 → 0.899, multi-label 0.453 → 0.789. Two gates that never learned to fire at all from scratch reach 0.857 and 0.583 recall from a handful of positives.

Licence

Apache 2.0.