ODIN · T0 open

Open tier T0 — synthetics, weights, leaderboard

What makes ODIN special: an open synthetic cohort (CC0), model weights with a card, and a BYOM leaderboard. Everything downloads and reproduces right here.

Synthetic cohort (CC0) CC0

1000 synthetic records mirroring the marginal distributions of the real cohort. For validating pipelines, NOT for measuring accuracy (the joint structure is simplified).

Download · CSV

Model weights + card

The published RF ensemble (frozen-test AUROC 0.8213, n=747). Requires sklearn 1.6.1. Calibrated probabilities are flat — the live CDSS shows a percentile vs the cohort, not “risk N%”.

model-v0.pkl Model card → Features → Score percentiles →

Reproduce the baseline

baseline.py prints the ensemble AUROC from weights + synthetics. Expect below 0.82 — that is correct (synthetic joints are simplified).

pip install scikit-learn==1.6.1 pandas numpy
python3 baseline.py model-v0.pkl \
  synthetic-dental-implants-v0.csv

baseline.py →

Leaderboard — BYOM v0

Metric: AUROC

Baseline: AUROC 0.8213 (frozen) · 0.4677 (synthetic)

Held-out: real registry test split (frozen, n=113, Taiwan cohort) — NOT the synthetic release

No external submissions yet. The protocol is live — be the first.

Protocol: BYOM v0: your predictor reads the synthetic CSV (same columns, WITHOUT unfavorable_outcome) on stdin and writes one probability [0,1] per row to stdout — any runtime (script or `docker run -i`). Self-check first: `node validate-submission.mjs -- <your command>` (script in this folder; fetches the cohort itself). Then email info@dentum.uz — the registry runs your submission on the frozen held-out set (n=113) manually during the pilot.

Submission self-check (one command):

curl -sO https://dentum.uz/registry/open/validate-submission.mjs && node validate-submission.mjs -- python3 predict.py

validate-submission.mjs → leaderboard.json →

Synthetics are CC0. Weights come from the research prototype (academic use; see the card). This is NOT a medical device.

API → · T0 stats →