What it was trained on
450 examples, one per case, from the practice split of the competition's
own cohort (cohort_dev.json). The targets are not human-written: they are
generated by a deterministic program that assembles the note the competition's
scoring function implies — verdict and confidence band in the question bank's own
wording, the five attributions in rank order with the direction each one gives,
values quoted only where the record can support them, and the findings the record
reports as present so that what it reports as absent follows by elimination.
So this is distillation of a program into a model, not a model taught medicine.
The held-out slice of 150 cases was excluded from training. The
competition's test split was never read.
Intended use
This model is for one task, in one prompt format, on synthetic data. It does not
diagnose, and it should not be pointed at real patient records. Every patient in
its training data is synthetic.
It expects the rendering the competition harness produces — the patient record,
the model output, the top 5 attributions and the citable-fields JSON, followed by
a short instruction — and replies with a single JSON object holding note and
drivers. Outside that format it will still answer, but nothing about its
behaviour there was measured.