vadimbelsky/qwen3.5-esi-triage-grpo-v49-merged

Qwen3.5-9B GRPO v49 — ESI Triage (Merged Full Model)

This is the merged full-weights version of the v49 GRPO ESI triage adapter — an iteration on v46 that targets specific clinical-rule failure modes identified in v46's error analysis.

What changed from v46

Error triage on v46's 8 wrong cases on MIETIC-36 revealed a clear pattern:

• 3 of 8 errors involved already-performed lifesaving interventions (intubation, central line, chest tube) that should trigger ESI-1 but were assigned ESI-2.
• 1 error missed the severe pain rule (pain ≥ 7 should anchor ESI ≤ 2).
• 1 error missed an open-fracture high-risk trigger.

v49 keeps everything that made v46 work (the 0.0 safety valve, asymmetric over-/under-triage penalties, 5× ESI-1 oversample) and adds rule-aware reward bonuses:

Trigger pattern in case text	Bonus / penalty
`intubat	chest tube
`intubat	chest tube
`open fracture	penetrating
Pain ≥ 7 + gold=2 + pred>2	-0.5

Other changes:

• Training budget raised from 512 → 1024 tokens (matches eval, eliminates the clipped-completions problem)
• No-parse penalty hardened: -0.5 → -2.0 (v48 lesson: the no-action baseline must be worse than every wrong action)
• Warm-start from v46 (not v43) — refining the best known policy
• 300 steps at LR=2e-7 (refinement, not relearning — avoids v47's hedging drift)

Training metrics (sanity)

v49 was the cleanest GRPO run of the series:

• clipped_ratio held at 4–19% throughout (vs v48's 90%+ for most of training)
• reward positive from step 10 (warm-start) with mid-run peak ~+0.49
• reward_std consistently 1.0+ — strong learning signal, no group collapse
• 300 steps in 17h48m on NVIDIA GB10

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

repo = "vadimbelsky/qwen3.5-esi-triage-grpo-v49-merged"
tokenizer = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, torch_dtype=torch.bfloat16, device_map="auto")

SYSTEM = (
    "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."
)

case = ("A 78-year-old female arrived intubated for airway protection after being found "
        "unresponsive. Central line placed. BP 120/58, HR 150, RR 20, SpO2 97%. "
        "Fever and presumed pneumonia. ")

prompt = tokenizer.apply_chat_template(
    [{"role": "system", "content": SYSTEM},
     {"role": "user",   "content": case}],
    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 behavior on the example above: the model should now recognize that an already-intubated patient with a central line meets ESI-1 Step A criteria (lifesaving intervention performed), and answer ANSWER: ESI 1 — fixing one of v46's known failure cases.

Evaluation

Evaluation results vs v46 will be added once the MIETIC-36 dual-mode eval completes. v46 baseline: 77.8% exact / 94.4% adjacent.

Limitations

Research model. Not approved for clinical use. Same caveats as v46. The rule bonuses target specific patterns we measured failing in v46 — they don't generalize to clinical rules outside that regex set. Cases that require rules outside {intubation, lifesaving intervention, open injury, severe pain} may still err.