Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,29 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
base_model: intfloat/multilingual-e5-small
|
| 4 |
+
library_name: transformers.js
|
| 5 |
+
tags: [sentence-transformers, feature-extraction, medical, on-device]
|
| 6 |
+
---
|
| 7 |
+
|
| 8 |
+
# evum/lab-marker-e5-small
|
| 9 |
+
|
| 10 |
+
A contrastive fine-tune of [multilingual-e5-small](https://huggingface.co/intfloat/multilingual-e5-small)
|
| 11 |
+
(MIT) for **lab-marker naming** in [Evum](https://github.com/) — it maps a lab-report label
|
| 12 |
+
(any language) to one of Evum's marker ids by embedding similarity. Runs **on-device** via
|
| 13 |
+
transformers.js; the user's health data never leaves the browser.
|
| 14 |
+
|
| 15 |
+
**Open weights, closed recipe:** the weights are public and the eval below is reproducible on
|
| 16 |
+
the open fixtures, but the training pipeline is proprietary.
|
| 17 |
+
|
| 18 |
+
## Eval (held-out fixtures, precision-first)
|
| 19 |
+
| model | precision | recall | threshold | margin |
|
| 20 |
+
|---|---|---|---|---|
|
| 21 |
+
| base multilingual-e5-small | 1.0 | 0.8235294117647058 | 0.9 | 0.01 |
|
| 22 |
+
| **this model (q8 ONNX)** | 1.0 | 0.9411764705882353 | 0.84 | 0.01 |
|
| 23 |
+
|
| 24 |
+
## Use
|
| 25 |
+
Feature-extraction; prefix report labels with `query: ` and marker names with `passage: `,
|
| 26 |
+
mean-pool, L2-normalize, cosine similarity. Names only — it never produces a measurement value.
|
| 27 |
+
|
| 28 |
+
## Limitations
|
| 29 |
+
Small fine-tune; a naming *fallback* behind a deterministic catalog. Not medical advice.
|