Qwen3.5-9B GRPO v46 — ESI Triage (Merged Full Model)

This is the merged full-weights version of the v46 GRPO ESI triage adapter.

The LoRA adapter weights from vadimbelsky/qwen3.5-esi-triage-grpo-v46 have been merged directly into the Qwen3.5-9B base model used during training. This guarantees the model produces v46's exact trained output format (EXTRACTION: → ESI ALGORITHM: → ANSWER: ESI N) without the format drift seen when mounting the LoRA on a different base.

When to use this vs. the LoRA adapter

• Use this (merged model) if you want exact behavior reproduction, faster inference (no adapter overhead), or easier deployment in environments without PEFT support.
• Use the LoRA adapter if you want lower storage (~230MB vs ~18GB) and already have the same base model loaded.

Performance

77.8% exact accuracy / 94.4% adjacent accuracy on the 36-case MIETIC expert-annotated evaluation set. See the adapter repository for full training methodology, reward function design, and the iteration journey (v45c → v46 → v47 → v48 lessons).

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

repo = "vadimbelsky/qwen3.5-esi-triage-grpo-v46-merged"
tokenizer = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, torch_dtype=torch.bfloat16, device_map="auto")

messages = [
    {"role": "system", "content":
        "You are an expert emergency triage nurse. "
        "Extract clinical fields, apply the ESI algorithm step by step, then state the ESI level. "
        "Be concise — stay under 150 words total."},
    {"role": "user", "content":
        "A 67-year-old male arrived via ambulance with sudden onset chest pain "
        "radiating to the left arm, diaphoresis, and shortness of breath. "
        "BP 88/60, HR 118, RR 24, SpO2 91%. History of MI and hypertension. Pain 9/10."},
]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
out = model.generate(
    **tokenizer(prompt, return_tensors="pt").to(model.device),
    max_new_tokens=1024, temperature=0.1, do_sample=True,
)
print(tokenizer.decode(out[0], skip_special_tokens=True))

Expected output format:

EXTRACTION:
- Chief complaint: ...
- Vital signs: ...

ESI ALGORITHM:
- Step A: ...
- Step B: ...

ANSWER: ESI 1

Limitations

This is a research model. Not approved for clinical use. See the adapter repository for known weaknesses (e.g. occasional missed clinical rules around already-performed lifesaving interventions, severe pain, and open injuries).