Modelpublic

rishabhhota/medgemma-s2s3-combined-covid

sourceHugging Facegemmaupdated 7mo agoView on Hugging Face

0likes12downloads

Model Card

MedGemma S2+S3 Combined COVID-19 Assessment

Trained with Strategy 2+3 combined: synthetic multimodal pairing + single-modality QA. Single unified training phase with mixed data — no catastrophic forgetting.

Results

Task	Test Accuracy
Vitals risk (3-class)	100.0%
Symptoms COVID +/-	98.2%
Combined assessment	qualitative

Features

—Handles missing modalities (vitals-only, symptoms-only, or combined)
—Vitals augmentation (5x for minority class)
—Balanced symptom sampling (2000/class cap)
—5238 mixed training examples

How to load

python

from transformers import AutoModelForImageTextToText, AutoProcessor
from peft import PeftModel
import torch

base = AutoModelForImageTextToText.from_pretrained(
    'google/medgemma-4b-it', torch_dtype=torch.bfloat16,
    attn_implementation='flash_attention_2', device_map='auto')
cxr = PeftModel.from_pretrained(base, 'rishabhhota/medgemma-covid-cxr-finetuned')
merged = cxr.merge_and_unload()
model = PeftModel.from_pretrained(merged, 'rishabhhota/medgemma-s2s3-combined-covid')
model.eval()