Ephraimmm

medgemma-soap-finetuned1

README

License: apache-2.0

Overview

This model is a fine-tuned variant of MedGemma (Google's medical-domain Gemma 3 model family), adapted to draft clinical SOAP notes (Subjective, Objective, Assessment, Plan) from consultation text. It was built on top of unsloth/medgemma-4b-it-unsloth-bnb-4bit, a 4-bit quantized build of google/medgemma-4b-it, and fine-tuned with Unsloth and Hugging Face's TRL library. The goal is to help reduce the clinical documentation burden by structuring free-text encounter notes or transcripts into a standard SOAP format — as a research/portfolio project, not a deployed clinical tool.

Training Details

The values below are as reported by the model author in the original repository card. No trainer_state.json, evaluation logs, or adapter_config.json are included in the repository, so these figures could not be independently re-derived from raw training artifacts — they are presented here as author-reported metadata rather than independently verified benchmarks.

Table with columns: Setting, Value
Setting	Value
Base model	`unsloth/medgemma-4b-it-unsloth-bnb-4bit` (Unsloth's 4-bit build of `google/medgemma-4b-it`)
Architecture	Gemma3ForConditionalGeneration (Gemma 3 text backbone + SigLIP vision encoder), confirmed in `config.json`
Framework	Unsloth + Hugging Face TRL
Fine-tuning method	4-bit QLoRA
LoRA rank	16 (author-reported)
LoRA alpha	32 (author-reported)
Epochs	3 (author-reported)
Effective batch size	8 (author-reported)
Learning rate	2e-4 (author-reported)
LR scheduler	Cosine (author-reported)
Final train loss	0.8076 (author-reported)
Final eval loss	1.0012 (author-reported)
License	Apache 2.0
Language	English

This repository hosts the fully merged model weights (model-00001-of-00002.safetensors, model-00002-of-00002.safetensors) rather than a standalone LoRA adapter — the QLoRA weights described above were merged back into the base model before upload.

Intended Use

Drafting SOAP-formatted clinical notes from consultation text/transcripts, as a starting point for a clinician to review and edit.
Assisting with structuring unorganized clinical text into Subjective, Objective, Assessment, and Plan sections.
Educational and portfolio demonstration of applying parameter-efficient fine-tuning (QLoRA via Unsloth) to a medical-domain LLM.

This is a research/portfolio project and has not been evaluated or certified as a medical device. It is not intended for direct use in patient care without human clinical oversight, further validation, and regulatory review.

How to Use

The model uses the standard Gemma 3 chat template (<start_of_turn>user ... <end_of_turn>). Load it with transformers as follows:

python
import torch
from transformers import AutoProcessor, AutoModelForImageTextToText

model_id = "Ephraimmm/medgemma-soap-finetuned1"

processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

messages = [
    {
        "role": "user",
        "content": [
            {"type": "text", "text": (
                "Convert the following clinical consultation notes into a SOAP note "
                "(Subjective, Objective, Assessment, Plan):\n\n"
                "Patient reports a 3-day history of sore throat and low-grade fever. "
                "Temp 37.9C, throat erythematous, no exudate. Rapid strep negative. "
                "Plan: supportive care, follow up if symptoms worsen."
            )}
        ],
    }
]

inputs = processor.apply_chat_template(
    messages,
    add_generation_prompt=True,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
).to(model.device)

with torch.inference_mode():
    output = model.generate(**inputs, max_new_tokens=512)

response = processor.decode(output[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True)
print(response)

Note: the base architecture (Gemma3ForConditionalGeneration) supports multimodal image+text input, but this fine-tune was trained and is intended for text-only SOAP note generation.

Limitations

Not clinically validated. No formal clinical accuracy, safety, or bias evaluation has been performed or is included in this repository. Any accuracy or quality figures beyond the training/eval loss values above should not be assumed.
Not a substitute for professional medical judgment. Output must be reviewed and verified by a qualified clinician before any use in real patient documentation or care.
No independent training-run artifacts. Since trainer_state.json and adapter configuration files are not present in this repository, the exact LoRA/training hyperparameters could not be independently confirmed beyond what the author documented.
Domain and demographic coverage unknown. The fine-tuning dataset's size, sourcing, and representativeness are not documented in this repository, so generalization to arbitrary clinical specialties, populations, or documentation styles is unverified.
Hallucination risk. As with any generative LLM, the model may produce plausible-sounding but incorrect or fabricated clinical content (e.g., inventing findings not present in the source text).
Should not be used for real patient care, billing, or legal documentation without proper clinical validation, human review, and appropriate regulatory approval.

Author

Developed by Ephraimmm

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

Ephraimmm

Model Tree

Base

unsloth/medgemma-4b-it-unsloth-bnb-4bit

Fine-tuned

this model

Input Modalities

Text

Output Modalities