Modelpublic

ursvenuk/llama32-3b-soap-clinical

sourceHugging Facellama3updated 5mo agoView on Hugging Face

0likes37downloads

Model Card

Llama-3.2-3B SOAP Clinical NLP (Edge / Q4KM)

QLoRA fine-tuned on MTS-Dialog for FHIR-aligned SOAP note extraction from noisy ASR transcripts. Quantized to Q4KM GGUF for deployment on Raspberry Pi 5 and other edge devices.

Models

File	Epochs	Mean Field F1	Med Hallucination
llama323bsoap_q4km.gguf	1	0.725	11.8%
llama323bsoap3epochq4km.gguf	3	0.744	23.5%
llama323bsoap5epochq4km.gguf	5	0.751	39.4%

Recommended: epoch 1 model offers best clinical safety (lowest hallucination).

Usage (llama.cpp)

bash

llama-cli -m llama32_3b_soap_q4km.gguf -n 512 -c 4096 --temp 0.1 -p "<prompt>"

Pipeline

Audio → whisper.cpp (ASR) → llama.cpp (SOAP extraction) → FHIR-aligned JSON