This model is a fine-tuned version of Google Gemma-3-4B-IT optimized for medical image processing. It was developed as part of the MedGemma Impact Challenge competition using the reduced ROCO-v2 dataset.

📊 Evaluation Results

We evaluated the model using both rule-based technical metrics and GenAI-as-a-judge (Gemini 3 Flash).

Metric Name	Value
BERTScore F1	0.8468
BLEU	0.0185
ROUGE-L	0.1427

🚀 Model Description

Developed by: Aldabergenov Makhambet, Zhalgasbayev Arman
Language: English
Model Type: Multimodal Vision-Language Model (VLM)
Finetuned from: google/gemma-3-4b-it
Dataset: ROCO (10k balanced radiology images)

🛠 Usage Example

To use this model, you need the transformers and accelerate libraries.

python

from transformers import Gemma3ForConditionalGeneration, AutoProcessor
import torch

model_id = "Aldabergenov1/medgemma_roco_fine_tuned"
model = Gemma3ForConditionalGeneration.from_pretrained(
    model_id, 
    torch_dtype=torch.bfloat16, 
    device_map="auto"
)
processor = AutoProcessor.from_pretrained(model_id)

# Example Inference
# messages = [{"role": "user", "content": [{"type": "image"}, {"type": "text", "text": "Radiology report:"}]}]
# ... (rest of the inference code)