loll070303/one-apple-a-day-mlx

One Apple a Day MLX

This repository contains a lightweight LoRA adapter for mlx-community/Qwen3.5-4B-MLX-4bit, fine-tuned for local MLX inference and low-token medical triage-style responses.

The project was designed around a simple idea: instead of making the model larger, make the response policy more efficient. The adapter pushes the base model toward short, structured, actionable outputs that are easier to run locally, easier to parse, and better suited for edge-oriented workflows.

Project Focus

This model was created as a local-first MLX experiment:

• LoRA fine-tuning performed locally;
• inference tested locally;
• target hardware: MacBook with M4 Pro and 24 GB unified memory;
• framework: MLX / mlx-lm;
• target behavior: concise emergency/triage-style clinical responses.

The goal is not long-form medical reasoning. The goal is a compact assistant that can produce a rapid first-pass clinical response with a predictable schema and a much smaller generation budget.

Why This Is Edge-Oriented

This adapter is designed for edge deployment experiments on Apple-device workflows:

• it is a small LoRA adapter rather than a full model checkpoint;
• it uses a 4-bit MLX base model;
• it encourages short generations;
• it uses a rigid response schema;
• it avoids long explanatory chains unless needed;
• it is intended for local inference rather than cloud inference.

The adapter file is approximately 1.8 MB. The base model remains external and is loaded from:

mlx-community/Qwen3.5-4B-MLX-4bit

This makes the release lightweight while keeping it tied to a reproducible MLX base model.

Response Format

The model is trained to produce compact clinical triage cards:

**Top Presumed Diagnosis:** [specific diagnosis]
**Immediate Next Steps:**
- [specific action]
- [specific action]
- [specific action]

The intended behavior is:

• identify the likely high-priority diagnosis;
• avoid generic filler;
• name concrete tests, treatments, or immediate actions;
• stay brief;
• minimize generated tokens.

Dataset Construction

The fine-tuning dataset was created from samples originating in previously tested medical datasets. These samples were rewritten/summarized into a standardized response format, then reviewed and corrected before training.

Each sample was shaped toward:

• a concise diagnostic reasoning style;
• a single top presumed diagnosis;
• a short list of immediate next steps;
• reduced verbosity;
• consistent formatting.

The final training format used chat-style messages with prompt masking, so the model was optimized mainly on the assistant response rather than on reproducing the system prompt or user case.

Training Setup

Training was performed locally with MLX / mlx-lm.

Key configuration:

model: mlx-community/Qwen3.5-4B-MLX-4bit
fine_tune_type: lora
data: data/manual_reviewed_messages
mask_prompt: true
iters: 1400
batch_size: 1
grad_accumulation_steps: 4
learning_rate: 0.000015
max_seq_length: 576
grad_checkpoint: true
lora_parameters:
  keys: ["self_attn.q_proj", "self_attn.v_proj"]
  rank: 8
  scale: 16.0
  dropout: 0.05

Hardware used:

MacBook with M4 Pro
24 GB unified memory
MLX / Apple Silicon local training

Benchmark Snapshot

Small internal benchmark: 6 emergency-style medical cases, comparing the base model with this LoRA adapter.

Generation limit:

max_tokens = 320

Measured output:

Metric	Base model	Fine-tuned adapter	Change
Average generation time	4.40s	1.88s	2.34x faster
Average output tokens	311.3	84.8	72.8% fewer tokens
Average output words	180.2	46.5	3.87x more compact

Interpretation:

The fine-tuned adapter does not make the base model inherently faster per token. Instead, it makes the model generate far fewer tokens while preserving a useful triage-style answer structure. This is the core optimization target of the project.

Example Behavior

Input:

A 34-year-old marathon runner collapses at mile 22. He is confused and seizing.
Na+ is 118 mEq/L. He had been drinking large amounts of water throughout the race.

Expected style:

**Top Presumed Diagnosis:** Severe symptomatic hyponatremic encephalopathy
**Immediate Next Steps:**
- Administer hypertonic saline immediately.
- Give 100 mL of 3% NaCl bolus over 10 minutes for seizure or coma.
- Monitor serum sodium closely and avoid overly rapid correction.

Loading with MLX

from huggingface_hub import snapshot_download
from mlx_lm import load, generate
from mlx_lm.sample_utils import make_logits_processors, make_sampler

adapter_path = snapshot_download("loll070303/one-apple-a-day-mlx")

model, tokenizer = load(
    "mlx-community/Qwen3.5-4B-MLX-4bit",
    adapter_path=adapter_path,
)

prompt = """<|im_start|>system
You are an ER physician. Provide rapid, accurate, and highly specific clinical decisions.
RULES:
1. Objective data overrides subjective narrative.
2. Do not hallucinate.
3. NO GENERIC FILLER: name specific tests and medications.
4. Outside reasoning tags, output only diagnosis and immediate next steps.<|im_end|>
<|im_start|>user
A 45-year-old male with alcohol use disorder presents with confusion, ataxia, and ophthalmoplegia.<|im_end|>
<|im_start|>assistant
"""

sampler = make_sampler(temp=0.0)
logits_processors = make_logits_processors(
    repetition_penalty=1.08,
    repetition_context_size=128,
)

response = generate(
    model=model,
    tokenizer=tokenizer,
    prompt=prompt,
    max_tokens=320,
    sampler=sampler,
    logits_processors=logits_processors,
)

print(response)

Recommended Use Cases

This adapter is best suited for:

• local-first MLX demos;
• edge-oriented medical AI prototypes;
• rapid triage-style summarization;
• structured clinical response generation;
• low-token inference experiments;
• Apple Silicon local AI workflows.

It is not optimized for:

• long medical explanations;
• exhaustive differential diagnosis;
• autonomous medical decision-making;
• replacing professional clinical judgment.

Files

• adapters.safetensors: final LoRA adapter weights
• adapter_config.json: MLX adapter configuration
• lora.yaml: training configuration

Important Medical Disclaimer

This is an experimental research model. It is not a medical device. It is not certified for clinical use. It can make incorrect, incomplete, or unsafe medical suggestions.

The model may:

• miss important diagnoses;
• omit urgent management steps;
• produce incorrect medication or dosing suggestions;
• over-compress complex cases;
• perform poorly on rare or ambiguous presentations.

Do not use this model for autonomous diagnosis, treatment, emergency care, or clinical decision-making without review by a qualified healthcare professional.

This release is intended for research, prototyping, local AI experimentation, and evaluation of low-token medical response behavior with MLX.