๐Ÿ DiabeticModels

Open diabetes models โ€” shown and hosted. Every build proves its math.

No black-box. Every model here ships with four things you can check yourself: its cited corpus, its recipe, its beat-base evaluation, and its receipt. A model isn't "proven" until it beats its base model on a held-out test โ€” and we publish that math, win or lose.

Models flow down to your devices and the on-device helpers. Your private health data never flows up. The facts behind every model are published openly at DiabeticTruth.

The build board

LocalDiabetic Foot-Care Specialist ๐Ÿ”จ IN BUILD

A small open model that organizes and explains diabetic foot care โ€” grounded in cited guidance, tuned to never diagnose. Destined to run on the LocalDiabetic edge box, on your own hardware.

1 ยท Corpus canary โœ“ 2 ยท Scale corpus 3 ยท Training pairs 4 ยท Cook 5 ยท Beat base or kill 6 ยท Host

Honest status: not yet cooked. The cited fact corpus is staged (20 facts from NIDDK & IWGDF 2023, every one sourced). No weights exist yet. When this model beats its base on a held-out foot-care eval, the receipt and the download land right here โ€” not before.

Download โ€” pending beat-base โœ•   See the facts it learns โ†’

Neuropathy & "Know Your Numbers" Specialists PLANNED

The next builds, once the foot-care cook proves the recipe. Same spine: cited corpus โ†’ cook โ†’ beat-base โ†’ host.

Facts already published: Neuropathy ยท A1C / targets.

How we prove a model

  1. Corpus, cited. Every fact traces to an authoritative source. No source โ†’ not in the corpus.
  2. Recipe, published. Base model, data, and settings โ€” reproducible.
  3. Beat base or kill. The tuned model must beat its base on a held-out, deterministic eval. If it doesn't, it doesn't ship โ€” and we say so.
  4. Receipt. The eval result is recorded honestly, win or lose, in the open cook ledger.
  5. Host. Only then does the model go up for download and flow down to your devices.