charakaweb/phi4-clinical-mlx

Phi-4-mini Clinical MLX (4-Bit Merged)

A specialized 3.8B biomedical & clinical reasoning model built on Microsoft's Phi-4-mini-instruct, optimized natively for Apple Silicon Metal acceleration via Apple MLX.

The model underwent a 3-stage transfer learning curriculum:

1. Stage 1 (STEM Foundation): 116,000 instruction pairs across NCERT Classes 6–12 (Physics, Chemistry, Biology) eliminating foundational science hallucinations.
2. Stage 2 (PubMed 2026 Evidence): 12 recent 2026 clinical update archives from NCBI FTP covering survival outcomes (OS, PFS, HR), targeted therapeutics, and clinical trial endpoints.
3. Stage 3 (Comprehensive Internal Medicine): Balanced multi-specialty clinical curriculum (cardiology, nephrology, endocrinology, pulmonology) with an active oncology replay buffer.

The LoRA adapter weights have been permanently fused into the base 4-bit weights (mlx_lm fuse) to deliver zero-latency execution.

---

📊 Benchmark Results (PubMedQA)

Evaluated on 50 biomedical research decision tasks from PubMedQA:

Model	Accuracy	Score	Avg Latency	Relative Improvement
Base Phi-4-mini (4-bit)	26.0%	13 / 50	1.02s / question	Baseline
Phi-4-mini Clinical MLX (Merged)	40.0%	20 / 50	0.93s / question	+53.8% relative gain

---

⚡ Quickstart with Apple MLX

Install mlx-lm:

pip install mlx-lm

CLI Generation

python -m mlx_lm.generate \
  --model <repo_id> \
  --prompt "<|user|>\nWhat are the first-line therapeutic recommendations for heart failure with preserved ejection fraction (HFpEF)?<|end|>\n<|assistant|>\n" \
  --max-tokens 512

Python API

from mlx_lm import load, generate

model, tokenizer = load("<repo_id>")
prompt = "<|user|>\nSummarize the mechanism of action of SGLT2 inhibitors in diabetic kidney disease.<|end|>\n<|assistant|>\n"
response = generate(model, tokenizer, prompt=prompt, max_tokens=300)
print(response)

Local OpenAI-Compatible Server

python -m mlx_lm.server --model <repo_id> --port 8080

---

⚕️ Clinical Disclaimer

This model is intended solely for biomedical research, educational exploration, and experimental evaluation. It is not an FDA-cleared medical device and must not be used as a substitute for professional clinical judgment, diagnosis, or treatment.

🏆 Verified Medical Benchmark Results

Benchmark	Scope	Tested Samples	Accuracy	Evaluation Hardware
PubMedQA	Clinical Trial Evidence Decisions	100	49.0%	Apple Silicon Metal GPU
MedQA (USMLE)	Medical Board Diagnostic Cases	100	53.0%	Apple Silicon Metal GPU