medgemma-v3-covid-cxr – AI Model by rishabhhota | AlphaNeural AI

You can deploy this model and start earning money today!

rishabhhota/
medgemma-v3-covid-cxr

like

0

flash-attention-2

google/medgemma-4b-it

Views

No views yet

MedGemmaV3 — COVID-19 CXR Classifier (Flash Attention 2 · Full bf16 LoRA · Focal Loss)

Test accuracy: 95.59% · Task: 4-class CXR (COVID / Normal / Lung Opacity / Viral Pneumonia)

Per-class accuracy

Class	Recall
COVID	98.34%
Normal	95.88%
Lung Opacity	92.90%
Viral Pneumonia	98.02%

Key improvements over V2

Flash Attention 2 (3-4x faster on H100, lower memory)
Full bf16 LoRA — no 4-bit quantization (better accuracy)
Focal Loss γ=2, class-weighted + label smoothing (fixes minority classes)
Linear warmup → cosine LR schedule
Training augmentation (RandomHFlip, Rotation, ColorJitter)

Load for classification

python

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

base = AutoModelForImageTextToText.from_pretrained(
    'google/medgemma-4b-it',
    dtype=torch.bfloat16,
    attn_implementation='flash_attention_2',
)
backbone = PeftModel.from_pretrained(base, 'rishabhhota/medgemma-v3-covid-cxr')
# Then load classification_head/head.pt using head_config.json

Report generation fine-tuning

This LoRA adapter specialises MedGemma for COVID CXR reasoning. For report generation, load this adapter, then:

Option A: backbone.add_adapter(new_lora_config, adapter_name='report_gen') — stacked adapters
Option B: backbone.merge_and_unload() then apply a fresh LoRA for generation