PD Mel Gemma4 E4B Reasoning LoRA

LoRA adapter that teaches Gemma 4 E4B to write step-by-step visual reasoning over Mel spectrograms of voice recordings, conditioned on the ground-truth class (healthy or parkinsons).

Built for Parkinson’s voice / speech-signal explainability (EDGE research). Teacher traces were generated with Gemma 4 31B; this student is the deployable ~4B-class Unsloth QLoRA fine-tune.

Quick facts


Base	`unsloth/gemma-4-e4b-it-unsloth-bnb-4bit`
Type	Vision–language LoRA (PEFT)
Task	Mel image + class label → reasoning trace
LoRA	r=32, α=32
Train	~1134 Mel samples, 1 epoch / 270 steps, loss ≈ 0.30
Framework	Unsloth + TRL SFT

Intended use

Research demos and papers on explainable PD vs healthy Mel analysis
Not a clinical diagnostic device — labels are assumed given; model explains why the spectrogram fits the label

Output format

{REASONING_START}
1. ...
2. ...
3. ...
{REASONING_END}

Example (conceptual)

Input: Mel PNG of a sustained vowel + text Class label: parkinsons
Output: Numbered visual cues (harmonic instability, breathiness texture, formant blur, etc.) inside the reasoning tags.

How to load

from unsloth import FastVisionModel
from PIL import Image

model, processor = FastVisionModel.from_pretrained(
    "HF_USER/pd-mel-gemma4-e4b-reasoning-lora",
    load_in_4bit=True,
)
FastVisionModel.for_inference(model)

Or load this folder locally after download.

Training data (summary)

Mel spectrogram images (healthy / parkinsons)
Assistant targets: reasoning-only traces from google/gemma-4-31B-it (4-bit teacher on Kaggle T4×2)

Limitations

Depends on correct class label in the prompt
Visual reasoning can be plausible but not ground-truth physiology
4-bit base + LoRA; not a full merged FP16 dump unless you merge yourself

Citation

EDGE Parkinson Mel reasoning fine-tune (Gemma 4 E4B Unsloth LoRA). Guide: https://unsloth.ai/docs/models/gemma-4/train

Downloads last month: -