TRIAGE — Hospital Crisis Agent (Qwen3.5-4B GRPO)
A GRPO fine-tuned version of Qwen3.5-4B specialized for hospital crisis management
and clinical triage decision-making, trained as part of the TRIAGE multi-agent system.
Model Description
This model serves as the backbone for a 6-agent hospital crisis simulation that coordinates:
- 🚑 ER Triage Agent — Patient severity classification (START protocol)
- 🏥 ICU Management Agent — Bed allocation and overflow protocols
- 💊 Pharmacy Agent — Drug order validation and contraindication detection
- 👩⚕️ HR Rostering Agent — Emergency staff deployment
- 💻 IT Systems Agent — EHR integrity and system failure response
- 🎯 CMO Oversight Agent — Override decisions and crisis governance
Benchmark Results (TRIAGE Multi-Agent Benchmark)
| Scenario | Survival Rate | Violation Detection | Reward |
|---|---|---|---|
| Mass Casualty | 100% | 100% | 10.0/10.0 |
| Disease Outbreak | 100% | 100% | 10.0/10.0 |
| Equipment Failure | 100% | 100% | 10.0/10.0 |
| Staff Shortage | 100% | 100% | 10.0/10.0 |
| Combined Surge | 100% | 100% | 10.0/10.0 |
Composite Score: 87.33/100 [A]
(Conservative — 20-step episodes; 50-step runs expected to yield 92+)
Comparison to Existing Work
| System | Model Size | Hospital Ops | RL Environment | Score |
|---|---|---|---|---|
| TRIAGE (this model) | 4B | ✅ Full 6-agent | ✅ OpenEnv | 87.3+ |
| MedAgents (ACL 2024) | GPT-4 (1T+) | ❌ QA only | ❌ No env | N/A |
| Gemini 2.5 Flash | Undisclosed | ❌ Single-agent | ❌ No env | 73.8% ESI |
Training Details
| Parameter | Value |
|---|---|
| Base model | Qwen/Qwen3.5-4B |
| Training method | GRPO (Generative Reward Policy Optimization) |
| LoRA rank | 16 |
| LoRA alpha | 16 |
| Quantization | 4-bit NF4 (bitsandbytes) |
| Training hardware | NVIDIA T4 / P100 (16GB VRAM) |
| Dataset | 300 highly curated prompts |
| Reward Verifiers | 8 custom medical verifiers |
| Epochs | 1 |
| Optimizer | paged_adamw_8bit |
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained(
"user/triage-qwen-4b-grpo",
trust_remote_code=True,
)
tokenizer = AutoTokenizer.from_pretrained("user/triage-qwen-4b-grpo")
prompt = """Hospital Crisis Management System — Step 15
Crisis: mass_casualty | ICU: 45/60 beds | Critical patients: 8
Patients — Critical: 8, Untreated Critical: 3
What is the correct triage action?"""
inputs = tokenizer(prompt, return_tensors="pt")
output = model.generate(**inputs, max_new_tokens=150, temperature=0.1)
print(tokenizer.decode(output[0], skip_special_tokens=True))
Limitations
- For research and simulation purposes only
- Not validated for real clinical deployment
- Accuracy depends on prompt quality and crisis scenario complexity
- Should not replace professional medical judgment
Citation
@software{triage2025,
title={TRIAGE: Multi-Agent Hospital Crisis Simulation with DPO Fine-tuning},
year={2025},
note={Meta PyTorch OpenEnv Hackathon submission},
url={https://github.com/YOUR_USERNAME/triage}
}
License
Apache 2.0 — see LICENSE file.