You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this model content.

medgemma-4b_finetuned_175 (LoRA)

LoRA adapter cho google/medgemma-4b-it, fine-tune sinh Impression (kết luận X-quang tiếng Việt) — dữ liệu nội bộ BV175 (đa bộ phận).

  • Task: Impression Generation (Stage 2).
  • Loss: cross-entropy (mask token assistant), Optim: AdamW, bf16.
  • LoRA: r=16, alpha=32, dropout=0.05, target = proj LLM (đóng băng vision SigLIP).
  • Data: 1210 train / 156 val / 157 test (split theo bệnh nhân).

Kết quả (test 157, impression)

ROUGE-1 ROUGE-2 ROUGE-L BLEU
zero-shot 26.8 9.4 17.8 0.9
fine-tuned 55.9 34.8 45.5 18.3

Dùng

from transformers import AutoProcessor, AutoModelForImageTextToText
from peft import PeftModel
base = "google/medgemma-4b-it"     # gated: cần chấp nhận license + đăng nhập HF
proc = AutoProcessor.from_pretrained(base)
model = AutoModelForImageTextToText.from_pretrained(base, torch_dtype="bfloat16", device_map="cuda")
model = PeftModel.from_pretrained(model, "<user>/medgemma-4b_finetuned_175")

Base model MedGemma là gated (phải xin quyền trên HF).

Downloads last month
5
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for nvtkienn/medgemma-4b_finetuned_175

Adapter
(123)
this model