Access the MedGemma caption adapters

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

These LoRA adapters are derivatives of google/medgemma-4b-it and are governed by the Gemma Terms of Use and the Health AI Developer Foundations terms. Research and evaluation only, not for clinical use.

Figure-captioning LoRA adapters (MedGemma-4B)

PEFT LoRA adapters (r=16, alpha=32) fine-tuned on medical figure captions on top of google/medgemma-4b-it. They accompany a SPIE Medical Imaging study on caption fine-tuning for medical figure captioning, and are released for research and evaluation only. Not for clinical use. Access is gated because these are Gemma derivatives (Gemma + Health AI Developer Foundations terms). Manuscript under submission.

Conditions (subfolders; seeds 42/43/44)

c1_raw_s{seed}: fine-tuned on raw figure captions.
c2_clean_s{seed}: fine-tuned on cleaned captions.
c3_aux_s{seed}: cleaned captions plus an auxiliary content loss.

C0 (zero-shot) is the base model with no adapter.

Loading

from peft import PeftModel
from transformers import AutoModelForImageTextToText
base = AutoModelForImageTextToText.from_pretrained("google/medgemma-4b-it", torch_dtype="bfloat16")
model = PeftModel.from_pretrained(base, "akmaier/medgemma-4b-figcaption-lora", subfolder="c2_clean_s42")

Notes

Trained on medical figure captions from open-access and text-and-data-mining sources (figures only; no captions or images are shipped). Findings: fine-tuning raises content overlap and hallucination together; the auxiliary loss does not help; for this 4B model caption cleaning does not reduce hallucination and instead trims content. The models do not reliably perceive imaging artifacts. For research and evaluation, not for clinical decisions.