Duan

Duan is a fine-tuned Laya System-1 decision engine for emergency and clinical decision tasks. It answers a question about a patient state in a single forward pass — no autoregressive decoding, no chain-of-thought — and returns a typed decision together with a calibrated probability distribution, a confidence estimate, and an abstention signal.

Architecture

Backbone encoder	jhu-clsp/mmBERT-base (ModernBERT, 768 hidden, 322 M params)
Head	2-layer Transformer encoder + typed decision scorer (Embedding(3, d) type bias + LayerNorm/Linear/GELU/Linear scorer)
Sequence budget	max_len = 1024, head_max_len = 256
Task format	[CLS] <type> instructions [SEP] [MASK] opt₀ … [MASK] optₙ [SEP] state [SEP]

Three question types, sharing one scorer — the only structural difference is a 3-row type embedding:

type	criteria	gold keys	answer
choice	{label: description}	same labels	argmax option + distribution
score	[level 0 desc, level 1 desc, …]	"0" … "n-1"	expected level + distribution
noul	{"false": …, "true": …}	false / true	P(true)

Output

Decision · Probability · Confidence · Abstention — a full distribution over options, an expected level for score, P(true) for noul, plus a calibrated confidence and an action head that can abstain rather than guess.

Training

Objective	RLCD — group sampling from perturbed logits → strictly proper scoring reward (log score + spherical, + RPS for score) → group-normalised advantage × policy gradient, plus soft-label cross-entropy
Steps	12,992 optimizer steps (8 epochs)
Effective batch	64 sequences/step
Data	103,886 train items + 1,500 held-out for calibration (stratified)
Hardware	1 × A800 80 GB, bf16, 5.2 h
Seeds	20260925 (train) / 20260922 (calibration split)

Training data (8 sources, 3 languages)

Source	Task	Lang	Licence
NHAMCS 2018–2022 (CDC/NCHS)	ED triage (ESI 1–5)	en	MIT (US public data)
UCI Diabetes 130-US hospitals	readmission window (score, 3 levels)	en	CC BY 4.0
MedQA — US / Mainland / Taiwan	medical MCQ	en / zh-Hans / zh-Hant	see upstream dataset
CARE-Bench	care escalation (4 levels)	en	CC BY-NC 4.0
DDXPlus (English)	differential-diagnosis re-ranking	en	CC BY 4.0
PubMedQA PQA-L	conclusion supported (noul)	en	MIT

⚠️ No source's raw text is redistributed here — only the resulting weights. Several sources are not clinical outcomes: MedQA labels are exam answer keys, CARE-Bench labels are human-annotated consensus grades, PubMedQA labels are literature annotations, and DDXPlus is fully synthetic.

Evaluation (test split, n = 10,538)

Task	n	acc	majority	macro-F1	ECE	other
ddx_diagnosis	2000	0.9905	0.1660	0.9900	0.0071	synthetic data
ed_triage (choice, 5)	2000	0.5600	0.4975	0.3490	0.1423	QWK 0.4290
readmission_window (score, 3)	2000	0.5810	0.5350	0.4178	0.0391	QWK 0.2239, MAE 0.4875, ±1 level 93.2 %
medqa_mcq	4415	0.3647	0.2258	0.3659	0.4564	en 0.2620 / zh-Hans 0.4505 / zh-Hant 0.3220 (acc)
care_escalation	79	0.6456	0.3418	0.6317	0.2951	small n
pubmedqa (noul)	44	0.7273	0.6364	0.6966	0.2727	AUROC 0.6708, PR-AUC 0.8135

Read these carefully. Majority-class baselines are given because every task is imbalanced; macro-F1 is the primary metric; cross-task macro-F1 is not comparable (different label spaces). ddx_diagnosis is high because DDXPlus is synthetic and its option set comes from the generator's own posterior — this is closed-set re-ranking, not open-set diagnosis.

Calibration

rl_agent_config.json ships a temperature_by_options table keyed by "<type>:<n_options>" and an additional temperature_by_qid table keyed by "<qid>|<type>:<n_options>". The per-qid table exists because the coarse bucket key cannot separate tasks that share a bucket — MedQA (5 options), NHAMCS triage (5 options) and CARE-Bench (4 options) all fall into choice:3-5, while their uncalibrated ECE differs by an order of magnitude.

To use it, rescale probabilities offline — the model emits p = softmax(logits / t), so p_new ∝ p^(t_old / t_new):

def rescale(probs, t_from, t_to):
    e = t_from / t_to
    w = {k: max(v, 1e-12) ** e for k, v in probs.items()}
    s = sum(w.values())
    return {k: v / s for k, v in w.items()}

Values are clamped to [0.5, 5] on load; any out-of-range fitted value is also recorded under temperature_by_qid_raw.

Usage

import laya

agent = laya.Agent("wipen/Duan")
res = agent.predict(
    state={"age": 67, "chief_complaint": "chest pain", "heart_rate": 118, "spo2": 91, ...},
    questions={"acuity": {
        "type": "choice",
        "instructions": "What ESI acuity level should this visit be triaged to?",
        "criteria": {"1": "requires immediate life-saving intervention", ...},
    }},
)
print(res["answers"]["acuity"]["choice"], res["answers"]["acuity"]["probability"])

Licence

Weights released under CC BY-NC 4.0, because one training source (CARE-Bench) is licensed CC BY-NC 4.0 and another (DDXPlus) requires attribution. The base model (convaiinnovations/laya-multilingual) and the training code are Apache-2.0.

Commercial use is not permitted under this licence. If you need commercial rights, the CARE-Bench portion must be removed and the model retrained, or separate permission obtained from that dataset's authors.

Citation

@misc{duan2026,
  title        = {Duan: A Healthcare System 1 Decision Model},
  author       = {Han, Weipeng},
  year         = {2026},
  howpublished = {\url{https://github.com/wipen/Duan}},
  note         = {Work in progress}
}

If you use the underlying engine, please also cite Laya:

@misc{laya,
  title        = {Laya: fast, non-autoregressive System 1 decision engine},
  author       = {{ConvAI Innovations}},
  howpublished = {\url{https://github.com/convaiinnovations/laya}},
  note         = {Apache-2.0}
}