seawer/clario-gemma-4-e4b-lora

Clario: Gemma 4 E4B LoRA for symptom-diary extraction

A LoRA adapter for google/gemma-4-E4B-it that reads a patient's diary note, written in plain English or Italian, and returns the symptoms in it as structured JSON: each term with the exact words it came from, mapped to the Human Phenotype Ontology (HPO), with negation, family history, severity and onset hints.

It is the patient-side model of Clario, a private symptom diary with a clinic review console, built for the lablab.ai AMD Developer Hackathon (Track 2, Health). The output is input for a clinician's review. It is not a diagnosis and must not be used as one.

Results

Entity F1 on the same 100 held-out synthetic diary rows from the test split of clario-synth-v3. "After resolver" applies Clario's ontology check (clario.extract.resolve_hpo) to every model's output before scoring.

Model	F1 after resolver	Raw F1
This adapter on Gemma 4 E4B (vLLM)	0.693	0.607
Gemma 4 31B IT (AMD Token Factory)	0.471	0.048
Gemma 4 E4B IT, no fine-tune	0.356	0.002

Result files with model, backend, date and n are in the repository under eval/results/.

Training

• Data: clario-synth-v3, 2,392 training rows of synthetic diaries written by Gemma 4 31B IT through the AMD Token Factory API, with gold labels by construction. No real patient data.
• Method: bf16 LoRA on the language model's attention and MLP layers (vision and audio towers untouched), r 16, alpha 16, dropout 0.05, learning rate 2e-4, 3 epochs. The checkpoint was chosen on validation F1 after resolver from a sweep over r 16/32 and lr 1e-4/2e-4.
• Hardware: one NVIDIA L40S on Modal.
• Prompt: the system prompt in train/prompt.py of the repository, with the diary as the only user turn. Use the same prompt at inference.

Use

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("google/gemma-4-E4B-it", torch_dtype="bfloat16", device_map="auto")
model = PeftModel.from_pretrained(base, "seawer/clario-gemma-4-e4b-lora")
tok = AutoTokenizer.from_pretrained("seawer/clario-gemma-4-e4b-lora")

For serving, merge the adapter into the base weights and run it with vLLM, as the Clario demo does.

Limits

• Trained and evaluated on synthetic diaries only. Real patient writing will differ.
• English and Italian only.
• Every extracted term needs a person's review; the Clario app sends unsure terms to the clinician.
• Use is subject to the Gemma terms of use.