You can deploy this model and start earning money today!

udaybhan10/
Diabetes-Expert-Llama-3.2-3B-MLX

mlx-community/Llama-3.2-3B-Instruct-4bit

finetune

mit

Views

No views yet

⚕️ Diabetes Expert Llama-3.2 (3B) - MLX

MEDICAL DISCLAIMER: This model is for educational and research purposes only. It is not a medical device and should not be used for clinical diagnosis or treatment decisions. Always consult a qualified healthcare provider.

📌 Model Description

The Diabetes Expert SLM is a high-precision language model fine-tuned for diabetes care and endocrinology. It is based on Llama-3.2-3B-Instruct and optimized for Apple Silicon using the MLX framework.

The model utilizes a "Triple-Threat" hybrid data strategy:

Local Distillation: Deep biology gleaned from PubMed via deepseek-r1 distillation.
2026 ADA Standards: Incorporates the latest Automated Insulin Delivery (AID) and Blood Pressure target guidelines.
Predictive Patterns: Grounded in historical clinical predictive datasets (Pima).

🚀 Key Features

Hallucination Suppression: Specifically trained to avoid common medical SLM errors like hemodynamic/glycemic term conflation.
Apple Silicon Native: Runs at ~84 tokens/sec on M4 Pro hardware.
Low Memory: Requires less than 3GB of Unified Memory.

🛠️ Usage (MLX)

You can load this model using the mlx-lm library on macOS:

python

from mlx_lm import load, generate

model_id = "mlx-community/Llama-3.2-3B-Instruct-4bit"
adapter_path = "udaybhan10/Diabetes-Expert-Llama-3.2-3B-MLX"

model, tokenizer = load(model_id, adapter_path=adapter_path)

prompt = "What is the molecular mechanism of beta-cell dedifferentiation in Type 2 Diabetes?"
response = generate(model, tokenizer, prompt=prompt, max_tokens=512)
print(response)

🧠 Training Details

Base Model: Llama-3.2-3B-Instruct-4bit
Method: QLoRA (MLX)
Iterations: 300
Learning Rate: 5e-5
LoRA Rank: 16
LoRA Alpha: 16
Hardware: M4 Pro (Unified Memory)

📊 Evaluation Results

Standard Benchmarks

Benchmark	Metric	Result
MedMCQA (Diabetes Subset)	Accuracy	83.33%
PubMedQA	Accuracy	52.00%

Training & Performance

Final Training Loss: 0.153
Inference Speed: ~84 tokens/sec (Apple M4 Pro)
Memory Footprint: < 3.0 GB Unified Memory

Qualitative Outcomes

Mechanistic Grounding: High proficiency in explaining molecular pathways (e.g., PDX1/MAFA depletion).
Guideline Compliance: 100% adherence to 2026 ADA revisions for AID and BP targets.
Hallucination Suppression: Successfully purged specific clinical term conflations.