README.md

---
base_model: vadimbelsky/qwen3.5-medical-ft-stage2
tags:
  - text-generation-inference
  - transformers
  - unsloth
  - qwen3_5
  - medical
  - triage
  - emergency-medicine
  - dpo
  - rlhf
  - medical-llm
  - clinical
  - healthcare
  - medicine
  - medical-ai
  - clinical-decision-support
license: apache-2.0
language:
- en
---

# Qwen3.5-9B Medical Triage — Stage 3 DPO (v4)

Emergency department triage model fine-tuned on Qwen3.5-9B via a 3-stage pipeline:
**Stage 1** (general medical SFT) → **Stage 2** (ED intake SOAP → ESI decision SFT) → **Stage 3** (DPO alignment to reduce over-triage, this model).

Quantized to **Q4_K_M GGUF** for on-device inference.

---

## Model Description

Given an ED SOAP intake note, the model outputs a structured triage decision:
- **ESI level** (1–5) with justification
- Key clinical findings
- Time-to-provider target
- Immediate interventions required

**ESI Scale:** 1 = Immediate life threat · 2 = Emergent high-risk · 3 = Urgent stable · 4 = Less urgent · 5 = Non-urgent

---

## Training Pipeline

| Stage | Method | Objective |
|-------|--------|-----------|
| 1 | SFT (LoRA r=16) | General medical knowledge (PubMed, clinical guidelines) |
| 2 | SFT (LoRA r=16) | SOAP note → structured ESI triage decision |
| 3 | DPO (LoRA r=8) | Reduce over-triage · preserve ESI 1/2 high-risk recall |

### Stage 3 DPO Details

- **Base:** Stage 2 LoRA checkpoint (`vadimbelsky/qwen3.5-medical-ft-stage2`)
- **Dataset:** `dpo_dataset_v4.jsonl` — 5,413 raw pairs → 7,789 weighted pairs
- **Loss:** Combined `apo_down × 0.3 + sft × 1.0` (MPO-style)
- **Beta:** 0.5 · **LR:** 5e-5 · **Epochs:** 0.1 (47 steps)
- **Batch:** 2 × 8 gradient accumulation = effective 16
- **ESI label prepending:** All chosen/rejected completions prefixed with explicit ESI label (e.g. `ESI 2 — Emergent (high risk)\n\n...`) to anchor preference signal at token position 0

### Dataset Sources (v4)

| Source | Description | Raw pairs | Weight | Weighted |
|--------|-------------|-----------|--------|---------|
| A | Anti-overtriage synthetic (ESI 3→1/2 rejected) | 2,388 | 1× | 2,388 |
| B | Anti-overtriage synthetic (ESI 4/5→1/2 rejected) | 1,500 | 1× | 1,500 |
| C | Edge cases (synthetic boundary scenarios) | 39 | 1× | 39 |
| D | ESI 1/2 anchor pairs (high-risk recall preservation) | 890 | 3× | 2,670 |
| E-over | ESI 3 bidirectional — anti-overtriage | 297 | 2× | 594 |
| E-under | ESI 3 bidirectional — anti-undertriage | 299 | 2× | 598 |
| **Total** | | **5,413** | | **7,789** |

---

## Evaluation Results

Evaluated on **MIMIC-IV-Ext Triage Instruction Corpus** (MIETIC) — 36 human-expert validated RETAIN cases.

### v4 vs Previous Stages

| Metric | Stage 2 (SFT) | v1 DPO | v2 DPO | v3 DPO | **v4 DPO** | Target |
|--------|--------------|--------|--------|--------|-----------|--------|
| Accuracy | ~68% | 55.6% | 50.0% | 27.8% | **75.0%** | >82% |
| Over-triage rate | ~22% | 22.2% | 30.6% | 0% | **13.9%** | <10% |
| Under-triage rate | ~8% | 36.1% | 41.7% | 72.2% | **11.1%** | <6% |
| High-risk recall (ESI 1+2) | ~84% | 76% | 64% | 40% | **92%** | 100% |
| ESI 3 accuracy | ~45% | ~40% | ~30% | ~0% | **60%** | >65% |

### v4 Detailed Results (MIETIC, n=36)

```
Samples evaluated   : 36
ESI level parsed    : 36 / 36
Correct             : 27
Accuracy            : 75.0%
Under-triage rate   : 11.1% (4 cases)
Over-triage rate    : 13.9% (5 cases)
High-risk recall    : 92.0% (ESI 1+2, n=25)
```

**Per-ESI Accuracy:**

| ESI Level | N  | Correct | Accuracy |
|-----------|----|---------|----------|
| ESI 1     | 14 | 12      | 85.7%    |
| ESI 2     | 11 | 9       | 81.8%    |
| ESI 3     | 5  | 3       | 60.0%    |
| ESI 4     | 4  | 2       | 50.0%    |
| ESI 5     | 2  | 1       | 50.0%    |

**Confusion Matrix** (rows = ground truth, cols = predicted):

```
GT \ Pred  ESI 1  ESI 2  ESI 3  ESI 4  ESI 5
ESI 1         12      2      0      0      0
ESI 2          0      9      2      0      0
ESI 3          0      2      3      0      0
ESI 4          0      0      2      2      0
ESI 5          0      0      0      1      1
```

All remaining errors are ±1 ESI boundary confusions — no catastrophic mis-triage.

---

## Key Lessons from DPO Iteration

- **v1–v3 failure:** IPO/sigmoid loss collapsed when dataset direction was 100% anti-overtriage → catastrophic under-triage regression (40% high-risk recall at worst)
- **v4 fix:** (1) ESI label prepended at token position 0 for unambiguous preference signal; (2) `apo_down + sft` combined loss preserves ESI 1/2 recall via SFT component; (3) Sources D (ESI 1/2 anchors ×3) + E (ESI 3 bidirectional ×2) balance dataset direction

---

## Usage

```python
# Requires llama.cpp server running with the Q4_K_M GGUF
# llama-server --model qwen3.5-medical-ft-stage3-dpo-q4km.gguf --port 8080 -c 4096

from openai import OpenAI
client = OpenAI(base_url="http://localhost:8080/v1", api_key="none")

SYSTEM_PROMPT = (
    "You are an expert emergency medicine triage nurse. "
    "Given a SOAP intake note, provide a structured triage decision including "
    "ESI level with justification, key clinical findings, time-to-provider target, "
    "and any immediate interventions required."
)

response = client.chat.completions.create(
    model="local",
    messages=[
        {"role": "system", "content": SYSTEM_PROMPT},
        {"role": "user", "content": "<SOAP intake note here>"},
    ],
    temperature=0.1,
    max_tokens=512,
)
print(response.choices[0].message.content)
```

---

## Limitations & Safety

> ⚠️ **This model is for research purposes only. It must NOT be used for clinical decision-making without licensed clinician oversight.**

- Evaluated on 36 MIETIC validation cases — not a clinical trial
- 11.1% under-triage rate means critical patients may be down-triaged
- 92% high-risk recall means ~8% of ESI 1/2 patients may be missed
- Model has not been validated on real ED populations
- Fine-tuned on synthetic + MIMIC-IV derived data only

---

## Training Infrastructure

- **Hardware:** NVIDIA GB10 (121 GB VRAM), 1 GPU
- **Framework:** Unsloth 2026.3.4 + TRL DPOTrainer + Transformers 5.2.0
- **Training time:** ~2 hours (47 steps)
- **Quantization:** GGUF Q4_K_M via llama.cpp

---

*Fine-tuned with [Unsloth](https://github.com/unslothai/unsloth) 🦥*