evum/lab-marker-e5-small

A contrastive fine-tune of multilingual-e5-small (MIT) for lab-marker naming in Evum — it maps a lab-report label (any language) to one of Evum's marker ids by embedding similarity. Runs on-device via transformers.js; the user's health data never leaves the browser.

Open weights, closed recipe: the weights are public and the eval below is reproducible on the open fixtures, but the training pipeline is proprietary.

Eval (held-out fixtures, precision-first)

model	precision	recall	threshold	margin
base multilingual-e5-small	1.0	0.8235294117647058	0.9	0.01
this model (q8 ONNX)	1.0	0.9411764705882353	0.84	0.01

Use

Feature-extraction; prefix report labels with query: and marker names with passage: , mean-pool, L2-normalize, cosine similarity. Names only — it never produces a measurement value.

Limitations

Small fine-tune; a naming fallback behind a deterministic catalog. Not medical advice.