Duan

Healthcare System 1

Fast decisions · Calibrated confidence · Edge deployment

Duan is a decision-native model for healthcare. Rather than generating long text, it returns a
typed decision, a probability distribution over the options, and a confidence value directly in a
single forward pass.

When confidence is insufficient, Duan is designed to hand the case to a human expert, or to a more
deliberative System 2.

---

Contents

1. Model base
2. Deployment and usage
3. Training method
4. Data
5. Evaluation
6. Capability boundaries
7. Citation
8. License

---

1. Model base

Duan is a fine-tune of Laya. Laya is an open-source non-autoregressive System 1 decision
engine that produces typed decisions over 100+ languages in a single forward pass.

CONVAI INNOVATIONS. Laya: fast, non-autoregressive System 1 decision engine. Apache-2.0.
https://github.com/convaiinnovations/laya

Laya ships three official checkpoints; Duan is built on the multilingual one:

Checkpoint	Encoder	Params	Token budget (max_len / head_max_len)	Languages
convaiinnovations/laya	ModernBERT-large	421M	512 / 192	English
convaiinnovations/laya-multilingual (Duan base)	jhu-clsp/mmBERT-base	322M	1024 / 256	100+
convaiinnovations/laya-typed-decisions	ModernBERT-large	421M	1024 / 256	typed-decision workflows

1.1 Architecture

Laya is a non-autoregressive architectural family with the following properties:

• Strictly proper scoring rules. The output distribution is trained not only with cross-entropy
  but against proper scoring rules, so the probability parameters can be read as probabilities.
• Typed questions + a single scorer. Every task is expressed as a question with a type —
  choice / score / noul — and the three types share the same scorer; the only structural
  difference is a 3-row type embedding (type_emb = Embedding(3, d)). There is no
  fixed-class-count classification head; the decision head applies a masked softmax over each
  sample's own option set.
• The option set is part of the input. The label space is rendered into the sequence at
  inference time, so the model is not restricted to a fixed list of decisions and can adapt flexibly
  to different decision scenarios.

Question type	criteria shape	Rendered as	Gold key space
choice	{label: description}	"<label>: <description>"	the labels
score	[level 0 description, level 1, …]	"level k: …"	"0" … "n-1"
noul	{false: …, true: …}	"false: no, it does not"	false / true

1.2 Sequence template and token budget

[CLS] <type> question: <instructions + rendered options> [SEP] [MASK] opt … [SEP] <state> [SEP]

The token budget is not a free parameter — it follows the checkpoint's own
rl_agent_config.json (1024 / 256 for this base). Measured on this corpus, the longest single
example is ≈366 tokens.

1.3 Output

{
  "decision": "review",
  "probabilities": { "routine": 0.08, "review": 0.84, "urgent": 0.08 },
  "confidence": 0.91,
  "abstain": false
}

The three question types return different payload keys. choice returns choice +
probabilities;

score returns score (an expected level) + probabilities, with no choice key;
noul returns only noul = P(true), with no probabilities key.

1.4 Download

The weights are released on two platforms (identical content); pick either:

Hugging Face — https://huggingface.co/wipen/Duan

pip install -U "huggingface_hub[cli]"
huggingface-cli download wipen/Duan

or in Python:

from huggingface_hub import snapshot_download
snapshot_download("wipen/Duan")

ModelScope — https://www.modelscope.cn/models/wipenhan/Duan

pip install -U modelscope
modelscope download --model wipenhan/Duan

or in Python:

from modelscope import snapshot_download
snapshot_download("wipenhan/Duan")

---

2. Deployment and usage

2.1 Requirements

Python	3.10 or newer
Hardware	CPU only — no GPU needed
Memory	~2 GB (322 M parameters, 643 MB of weights)
Network	first run downloads the weights (~643 MB), see §1.4

2.2 Agent deployment

Paste this one-line prompt into any coding agent that can run shell commands (Codex, WorkBuddy, Claude Code, Cursor, …) — it clones the repository, installs, starts and verifies everything:

Clone https://github.com/wipen/Duan and cd into it, then deploy Duan and start the local web demo: the weights come from https://huggingface.co/wipen/Duan (if huggingface.co is unreachable, download from https://www.modelscope.cn/models/wipenhan/Duan and point DUAN_MODEL at that local directory); run pip install -r requirements.txt, then python serve.py, then open http://127.0.0.1:8000 and confirm /health returns ok and one example runs.

The agent needs PyPI access (install) and, on first run, Hugging Face access (weights). The prompt above already names a ModelScope fallback; if HF is merely unreachable from the network, a mirror also works: export HF_ENDPOINT=https://hf-mirror.com.

When it finishes, open http://127.0.0.1:8000 to try it. Manual steps: §2.3.

2.3 Manual deployment

2.3.1 Install

From the repository root:

pip install -r requirements.txt

The file declares only laya[serve]; pip resolves it and installs torch, transformers, safetensors, huggingface_hub, fastapi and uvicorn as well.

2.3.2 Start the server

python serve.py                        # 127.0.0.1:8000
python serve.py --port 8080            # another port
python serve.py --model <id-or-path>   # another checkpoint (HF id or local dir)

Env var	Meaning	Default
DUAN_MODEL	HF repo id or local directory	wipen/Duan
DUAN_HOST / DUAN_PORT	bind address / port	127.0.0.1 / 8000
HF_ENDPOINT	Hugging Face mirror	—

The first start downloads and loads the weights; on CPU this takes about 20–30 s. It is ready when the log shows Uvicorn running on http://127.0.0.1:8000.

Method	Path	Purpose
GET	/	web demo (Playground)
GET	/health	health probe
POST	/predict	{"state": …, "questions": …} → decision
GET	/docs	interactive API reference

2.3.3 Try the demo

Open http://127.0.0.1:8000 in a browser and, in the Playground:

1. pick one case under Example use cases (real cases from the test split);
2. click Run;
3. you should get the decision, a probability bar per option, and the confidence. Switch EN / 中文 at the top right.

Or call the API directly:

curl -s http://127.0.0.1:8000/predict -H 'Content-Type: application/json' -d '{
  "state": {"age": 67, "chief_complaint": "chest pain", "heart_rate": 118},
  "questions": {"acuity": {"type": "choice",
    "instructions": "What ESI acuity level should this visit be triaged to?",
    "criteria": {"1": "immediate life-saving", "2": "high risk", "3": "many resources",
                 "4": "one resource", "5": "no resources"}}}
}'

Expected: answers.acuity.choice (the ESI level), answers.acuity.probabilities (per-level bars) and confidence.

2.3.4 Troubleshooting

Symptom	Cause / fix
ModuleNotFoundError: No module named 'laya'	dependencies missing → pip install -r requirements.txt
first start hangs on download	huggingface.co unreachable → export HF_ENDPOINT=https://hf-mirror.com
Address already in use	port taken → python serve.py --port 8080
page opens but Run errors	backend down or wrong address → start python serve.py; set the demo's Server box to the full URL, e.g. http://127.0.0.1:8000/predict
one decision takes ~0.4 s	normal on CPU; avoid firing many requests at once

---

3. Training method

Training is a full-parameter fine-tune of the multilingual base, distributed with DDP, driven by
Laya's RLCD objective.

3.1 The objective: RLCD

RLCD uses strictly proper scoring rules as the reward and a GRPO-style policy gradient.
Concretely:

Component	Value	Role
group_size	4	GRPO baseline sampling group
sigma_start → sigma_end	0.4 → 0.1	exploration noise, linearly annealed
w_sph	0.75	spherical score weight
w_rps	1.0	ranked probability score weight — applies to score questions only
ce_weight	1.0	soft-label cross-entropy weight, added to the policy-gradient term

The reward is a proper scoring rule rather than "points for answering correctly", which means the
model is optimised to recognise the uncertainty of its answer.

3.2 Optimization and batching

Setting	Value
epochs	8
micro_batch	8
target_effective_batch	64 sequences/step
max_tokens_per_batch	4096
lr_encoder	2.5e-5
lr_head	1.0e-4
weight_decay / min_lr / grad_clip	0.01 / 1.0e-6 / 1.0
precision	bf16 (auto: compute capability ≥ 8.0)
gradient / head checkpointing	on
freeze_encoder	false — full fine-tune
seed / calibration seed	20260925 / 20260922

3.3 Post-training temperature calibration

Temperature scaling happens after training, fitted on a held-out calibration slice, and it is
fitted per bucket rather than as one global value. The bucket name is determined by
(question type, option count). Merging data with mixed option counts under one temperature
concentrates the entire calibration error on whichever bucket is the minority.

3.4 Splits and anti-leakage

• The corpus is split into train and test.
• Leakage removal: 500 exact duplicates between MedQA-Mainland's official train and test splits
  were dropped from train (see §4.4).

---

4. Data

The corpus is 8 public sources across two languages (Chinese and English), converted uniformly
into Laya's typed-decision format.

4.1 Sources

#	Source	Language	Task	Options	train / test	Label nature	License
1	NHAMCS ED	en	ED triage (ESI 1–5)	5	20,000 / 2,000	observed clinical outcome	MIT
2	UCI Diabetes	en	30-day readmission window	3	20,000 / 2,000	observed clinical outcome	UCI ML Repo
3	MedQA (English)	en	medical knowledge single-choice	5	11,450 / 1,145	exam answer key	see §4.3
4	MedQA (Simplified Chinese)	zh-Hans	medical knowledge single-choice	5	20,000 / 2,000	exam answer key	see §4.3
5	MedQA (Traditional Chinese)	zh-Hant	medical knowledge single-choice	4	12,701 / 1,270	exam answer key	see §4.3
6	CARE-Bench	en	patient-facing triage and action recommendation	4	790 / 79	annotated expert consensus	CC-BY-NC-4.0
7	DDXPlus	en	automatic medical diagnosis and differential diagnosis	2–10	20,000 / 2,000	fully synthetic	CC-BY-4.0
8	PubMedQA (PQA-L)	en	conclusion supported (yes/no)	2	445 / 44	literature expert annotation	MIT
	Total				105,386 / 10,538

Sources 3–5, 7 and 8 are not clinical outcome data. MedQA labels are exam answer keys,
CARE-Bench labels are expert consensus grades, PubMedQA labels are literature annotations, and
DDXPlus labels come from a generative model. None of them supports any claim about clinical
prediction, incidence or causality. They are in the corpus so that the typed-decision machinery
is trained across multiple languages and option counts.

4.2 Shape of the corpus

Quantity	Value
train / test rows	105,386 / 10,538 (test = 10.00%)
train by language	en 72,685 (69.0%) · zh-Hans 20,000 (19.0%) · zh-Hant 12,701 (12.0%)
balancing	per-source equal weight, capped at 20,000 rows per source
option-count spread (train)	2: 1,435 · 3: 21,597 · 4: 14,920 · 5: 53,007 · 6: 1,586 · 7: 1,288 · 8: 1,556 · 9: 1,208 · 10: 8,789
corpus size on disk	train 122.8 MB · test 12.3 MB (jsonl)

4.3 Licensing

Model weights are licensed under CC-BY-NC-4.0, non-commercial use only; the code is Apache-2.0 (see §8).

Source	Licence	What it means here
NHAMCS ED	MIT	freely usable and redistributable
UCI Diabetes	UCI ML Repository terms	check terms before redistribution
PubMedQA	MIT	freely usable
DDXPlus	CC-BY-4.0	commercial use permitted with attribution
CARE-Bench	CC-BY-NC-4.0	non-commercial only
MedQA	per upstream dataset terms	see the source distribution

This repository does not ship the training corpus. The converted corpus is 135 MB of jsonl whose
rows contain verbatim source text. The corpus is therefore always regenerated locally from the
original sources by the conversion scripts, deterministically, with a fixed seed.

4.4 Known data defects we corrected

Source	Defect found	Handling
MedQA Mainland	genuine cross-split leakage: 221 test rows are exact duplicates of train rows (6.45%); 32 rows in dev∩test; a further 910 duplicates within train	leakage removed from train (500 rows) plus 1,140 within-batch duplicates
MedQA Taiwan	9 rows overlap with the held-out set	dropped
DDXPlus	the correct pathology sits at position 0 of the differential list 73.7% of the time	options deterministically shuffled by content hash; first-position hit rate falls to 17.3%, consistent with the uniform expectation
DDXPlus	icd10 and severity are deterministic functions of the label	excluded from the input state, kept only in metadata
PubMedQA	ships its own baseline model predictions (reasoning_free_pred, reasoning_required_pred)	asserted out of the input state
CARE-Bench	information_state is 100% pure for one class; round_role reaches 84% purity	removed from the input state, purity printed as evidence
NHAMCS ED	post-triage interventions, wait time, and vitals-missingness are all label proxies	excluded from the input state

---

5. Evaluation

Metrics are chosen from the label distributions rather than from convention — accuracy alone is
uninformative here: on NHAMCS triage the majority class already reaches 49.8% on this test
split.

The numbers below are measured on the test split (10,538 rows) after an 8-epoch full fine-tune.

Task	n	Accuracy	Majority baseline	Macro-F1	ECE
ed_triage (ESI 1–5)	2,000	0.5600	0.4975	0.3490	0.1423
readmission_window (3 levels)	2,000	0.5810	0.5350	0.4178	0.0391
ddx_diagnosis (2–10 options)	2,000	0.9905	0.1660	0.9900	0.0071
medqa_mcq (5 / 5 / 4 options)	4,415	0.3647	0.2258	0.3659	0.4564
care_escalation (4 levels)	79	0.6456	0.3418	0.6317	0.2951
pubmedqa (yes/no)	44	0.7273	0.6364	0.6966	0.2727

---

6. Capability boundaries

1. No causal inference. Statistical-association and spontaneous-reporting sources cannot
   support "drug X causes event Y", and no source of that class is used for any decision claim in
   this project.
2. No incidence estimation.
3. No clinical performance claim yet. All measured numbers are listed in §5,
   with their denominators and limitations. The 0.50–0.65 range often cited for 5-class ESI
   prediction is a literature expectation, not a result of this project.
4. No substitute for clinical judgement. Recall on high-risk minority classes is inherently
   limited; cases of that kind should be routed to a human via a confidence threshold.
5. No relicensing of third-party data. See §4.3; CARE-Bench is non-commercial
   and that constraint propagates to models trained on it.

---

7. 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, cite Laya as well:

@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}
}

---

8. License

Code: Apache-2.0 (see LICENSE); weights: CC-BY-NC-4.0, non-commercial use only (see LICENSE-MODEL).

Dataset licences are independent of this repository — see §4.3.

---

Safety

Duan is intended for research and clinical decision-support development. It is not a substitute
for qualified healthcare professionals, and must not be used for unvalidated autonomous diagnosis,
prescribing, or other high-risk medical decisions.