Add model card

README.md

@@ -0,0 +1,105 @@
+---
+base_model: unsloth/Qwen3.5-9B
+library_name: transformers
+pipeline_tag: text-generation
+tags:
+- medical
+- emergency-medicine
+- triage
+- dpo
+- rlhf
+- qwen3_5
+- unsloth
+language:
+- en
+license: apache-2.0
+---
+
+# Qwen3.5-9B — Medical Triage Stage 3 DPO
+
+Fine-tuned **Qwen3.5-9B** for emergency department triage decision support. This is the fully merged Stage 3 checkpoint, aligned with DPO to reduce over-triage and preserve recall for high-acuity patients.
+
+## Model Description
+
+| Property | Value |
+|---|---|
+| Base model | `unsloth/Qwen3.5-9B` |
+| Architecture | Qwen 3.5 — 9B parameters |
+| Precision | bfloat16 |
+| Training stages | 3 (SFT stage 1 → SFT stage 2 → DPO stage 3) |
+| Task | Emergency triage — ESI level assignment from SOAP notes |
+
+## Training — Stage 3 (this checkpoint)
+
+Stage 3 applies Direct Preference Optimization (DPO) on top of Stage 2 SFT adapters to correct over-triage bias.
+
+| Hyperparameter | Value |
+|---|---|
+| Method | DPO (sigmoid loss) |
+| Beta | 0.1 |
+| Learning rate | 5e-5 |
+| Epochs | 1 |
+| Batch size | 2 × 8 grad accum = 16 effective |
+| Optimizer | AdamW 8-bit |
+| LoRA rank | 16 |
+| Training pairs | 3,730 |
+| Validation pairs | 197 |
+| Framework | Unsloth + TRL |
+
+**Alignment objective:** reduce over-triage (assigning ESI 1/2 to ESI 3/4/5 cases) while preserving recall for truly high-acuity patients.
+
+## 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.
+```
+
+## Usage
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+import torch
+
+model_id = "vadimbelsky/qwen3.5-medical-ft-stage3-dpo"
+
+tokenizer = AutoTokenizer.from_pretrained(model_id)
+model = AutoModelForCausalLM.from_pretrained(
+    model_id,
+    dtype=torch.bfloat16,
+    device_map="auto",
+)
+
+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."
+)
+
+soap_note = """
+S: 67-year-old male, chest pain radiating to left arm x 45 minutes, diaphoretic, nausea.
+O: BP 160/95, HR 102, RR 20, SpO2 96% RA, temp 37.1°C.
+A: Suspected ACS / STEMI rule-out.
+P: 12-lead ECG, troponin, aspirin 325mg, IV access.
+"""
+
+messages = [
+    {"role": "system", "content": system_prompt},
+    {"role": "user", "content": soap_note.strip()},
+]
+
+inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True).to(model.device)
+outputs = model.generate(inputs, max_new_tokens=512, temperature=0.1, do_sample=True)
+print(tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=True))
+```
+
+## Intended Use
+
+- Clinical decision support tool for emergency triage education and research
+- **Not validated for clinical deployment** — always defer to licensed clinical staff
+
+## Limitations
+
+- Trained on a curated DPO dataset; real-world distribution may differ
+- Does not replace clinical judgment or access to vital signs, lab results, or imaging
+- Performance on rare ESI 1 presentations (e.g. cardiac arrest, airway emergency) should be independently validated