QLoRA adapter for unsloth/gemma-4-E2B-it, fine-tuned for conservative oral mucosal screening (binary categories: low_risk_or_variation, refer_for_clinical_review). Outputs structured JSON with recommendation and disclaimer; not a diagnostic model.

Load adapter

python

from unsloth import FastVisionModel
from peft import PeftModel

base_model, processor = FastVisionModel.from_pretrained(
    "unsloth/gemma-4-E2B-it",
    load_in_4bit=True,
)
model = PeftModel.from_pretrained(base_model, "YOUR_HF_USERNAME/no-vision-lora-15ep")

Training

Framework: Unsloth + TRL SFT, 15 epochs, QLoRA (r=32, alpha=64)
Dataset: cropped oral lesion images with instruction-following JSON labels (see project oral_gemma_finetune_package)

Notes

Requires the base model on Hugging Face; this repo contains adapter weights only (~228 MB).
For on-device deployment, merge and convert to LiteRT separately; validate runtime before production use.

Sidharth1743/orcal-cancer-gemma-4-e2b-finetuned

no-vision-lora-15ep

Load adapter

Training

Notes

Sidharth1743/
orcal-cancer-gemma-4-e2b-finetuned