You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Sona Laya (v2)

Fine-tune of convaiinnovations/laya-multilingual (322M, Apache-2.0) that picks which clinical card answers a clinician's request in Sona, including Manglish and Malay/English/Mandarin code-switching.

Question it was trained on: "Which clinical result answers this request?" over briefing, reasoning, xray, medication, prescription, followup, note, evidence or not-covered, plus "Which catalogue medicine does the request name?". The exact wording lives in Sona's packages/api/src/laya.ts; changing it changes the question.
Score: 46/50 (92%) on the card and 50/50 on the medicine, on a held-out set of 50 fictional requests; about 40 ms per request on Apple MPS. Untrained base model: 34%.
Data: about 1,600 fictional requests written by Gemini for two fictional demo patients. No real patient data.
Training: Laya's official Kaggle notebook, 2x T4, 4 epochs, about 6 minutes.

import laya
agent = laya.load("zhiyuantechnology/sona-laya")

Downloads last month: -

Safetensors

Model size

0.3B params

Tensor type

F16

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for zhiyuantechnology/sona-router

Base model

convaiinnovations/laya-multilingual

Finetuned

(77)

this model