Modelpublic

Phiphi216/medgemma-oncology-lora

sourceHugging Faceapache-2.0updated 4mo agoView on Hugging Face

0likes

Model Card

MedGemma Oncology Agent

This repository contains a LoRA fine-tuned oncology agent built on top of MedGemma 1.5 4B IT.

The model is trained to analyze chest CT scan slices together with structured clinical records and predict six-year lung cancer risk.

The model was fine-tuned using Unsloth and PEFT LoRA on a custom oncology dataset.

Model Details

Item	Value
Base Model	unsloth/medgemma-1.5-4b-it
Domain	Oncology
Task	Lung Cancer Risk Prediction
Finetuning Method	LoRA
Developer	Phiphi216
License	Apache 2.0

Dataset

Dataset:

—Phiphi216/oncology

Number of samples:

—2,000 training examples

Each sample contains:

—CT scan identifier
—Patient identifier
—Prompt
—Response

Response format:

json

{
  "lung_cancer_risk": "Cancer within follow-up"
}

Training Procedure

Base Model

markdown

unsloth/medgemma-1.5-4b-it

LoRA Configuration

yaml

r: 16
lora_alpha: 16
lora_dropout: 0.0

finetune_vision_layers: false
finetune_language_layers: true

finetune_attention_modules: true
finetune_mlp_modules: true

optimizer: adamw_torch_fused

learning_rate: 2e-4

weight_decay: 1e-3

lr_scheduler_type: cosine

max_grad_norm: 0.3

per_device_train_batch_size: 2

gradient_accumulation_steps: 2

effective_batch_size: 4

max_seq_length: 25808

max_steps: 125

warmup_ratio: 0