Fine-tuned from microsoft/Phi-3-mini-4k-instruct (3.8B parameters) using QLoRA on a medical triage instruction dataset. Given a patient symptom description, the model classifies urgency and provides a brief clinical rationale.

⚠️ Educational and portfolio use only. Not for clinical deployment.

Model Performance

Evaluated on 100 held-out synthetic patient presentations (25 per class).

Model	Triage Accuracy	ROUGE-L	EMERGENCY Recall
This model (fine-tuned)	64%	0.42	100%
Base Phi-3-mini (zero-shot)	37%	0.21	100%

The model never misses a genuine life-threatening EMERGENCY case.

Per-Class Accuracy

Level	Meaning	Accuracy
EMERGENCY	Call 999 immediately	100%
SEMI-URGENT	GP within 24–48 hours	60%
NON-URGENT	Routine GP appointment	92%
URGENT	Same-day care	4%

The URGENT class is frequently over-triaged to EMERGENCY — a conservative error that is clinically safer than the reverse.

Training Details

Property	Value
Base model	microsoft/Phi-3-mini-4k-instruct
Fine-tuning method	QLoRA — 4-bit NF4 + LoRA
LoRA rank (r)	16
LoRA target modules	qkv_proj, o_proj
Training examples	200 synthetic patient presentations
Training platform	Kaggle T4 GPU (free tier, 16GB VRAM)
Trainable parameters	9,437,184 (0.468% of total)
Training time	~8 minutes

How to Use

python

from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
import torch

model_id = "Milon96/phi3-medical-triage"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    torch_dtype=torch.float16,
    trust_remote_code=True,
)
pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)

instruction = """You are a medical triage assistant. Classify the following
patient presentation into: EMERGENCY, URGENT, SEMI-URGENT, NON-URGENT.
EMERGENCY: Life-threatening — call 999 immediately.
URGENT: Serious — same-day care required.
SEMI-URGENT: Needs care within 24-48 hours.
NON-URGENT: Routine — GP appointment within a week."""

patient = "72-year-old male with sudden chest pain radiating to his left arm, sweating and nausea. Started 15 minutes ago."

prompt = f"<|system|>\n{instruction}<|end|>\n<|user|>\n{patient}<|end|>\n<|assistant|>\n"
result = pipe(prompt, max_new_tokens=150, do_sample=False, temperature=1.0)
response = result[0]["generated_text"].split("<|assistant|>")[-1].replace("<|end|>","").strip()
print(response)
# Triage Level: EMERGENCY
# Symptoms are consistent with acute myocardial infarction (heart attack).
# Immediate emergency services required. Call 999 now.

Training Data

200 synthetic patient presentations generated using Claude (Anthropic) covering realistic clinical scenarios. Each example includes patient age, gender, and symptom description in natural language.

Key lesson from this project: MedQuAD (47,000 NIH Q&A pairs) uses a "What is X?" question format that does not match real patient presentations. 200 high-quality synthetic examples outperformed 1,500 MedQuAD examples — data quality beats data quantity.

Limitations

Trained on synthetic data — not validated on real clinical cases
URGENT class has low accuracy (4%) — frequently over-triaged to EMERGENCY
English language only
Not suitable for real medical decision-making

Repository

Full training code, dataset pipeline, and evaluation scripts: github.com/Milonahmed96/medical-triage-llm

Citation

bibtex

@misc{phi3-medical-triage,
  author = {Milon Ahmed},
  title = {Phi-3-mini Medical Triage Assistant},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/Milon96/phi3-medical-triage}
}