Instructions to use iatrode/iatro-clintrace with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use iatrode/iatro-clintrace with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="iatrode/iatro-clintrace")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("iatrode/iatro-clintrace")
model = AutoModelForCausalLM.from_pretrained("iatrode/iatro-clintrace", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use iatrode/iatro-clintrace with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "iatrode/iatro-clintrace"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "iatrode/iatro-clintrace",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/iatrode/iatro-clintrace

SGLang

How to use iatrode/iatro-clintrace with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "iatrode/iatro-clintrace" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "iatrode/iatro-clintrace",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "iatrode/iatro-clintrace" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "iatrode/iatro-clintrace",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use iatrode/iatro-clintrace with Docker Model Runner:
```
docker model run hf.co/iatrode/iatro-clintrace
```

You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

ClinTrace

ClinTrace is a research model that learns institutionally accumulated clinical decision experience from longitudinal records. Given the preceding course and current clinical evidence, it ranks reviewable plan--action candidates and associates each candidate with a profile of documented short-term clinical feedback. This release is the frozen ClinTrace v1.0.3 model developed and evaluated predominantly in hepatocellular carcinoma (HCC).

Research use only. ClinTrace is not a medical device and must not be used for autonomous diagnosis, triage, treatment selection or patient management.

Access

The complete weights are openly released under CC BY 4.0 through an automatically approved access gate on both hosted mirrors. The gate provides a responsible-use acknowledgement and does not involve case-by-case approval. After signing in and accepting the repository terms, the complete release can be downloaded immediately.

Hugging Face: https://huggingface.co/iatrode/iatro-clintrace
ModelScope: https://modelscope.cn/models/iatrode/iatro-clintrace
Versioned software archive: https://doi.org/10.5281/zenodo.21470033
Source code: https://github.com/iatrode/iatro-clintrace
Python package: https://pypi.org/project/iatro-clintrace/

The two model mirrors contain the same release artifact and manifest.

Installation and download

pip install iatro-clintrace

# Authenticate with either hosted mirror once.
hf auth login
# or: modelscope login

clintrace download
clintrace infer --help

clintrace download selects ModelScope for a China public IP and Hugging Face otherwise, reuses the corresponding local login, downloads the complete bundle and verifies that its manifest is present. The official clients can also download either repository directly; pass that local directory to clintrace infer --release-dir ....

Model contract

ClinTrace models a clinical decision point as:

ordered preceding-admission documents H + current clinical evidence E
    -> ranked plan--action (PX) candidate slots
    -> candidate-conditioned observed outcome (O) profiles

Evidence--Plan--eXecution--Outcome (EPXO) chains, timeline-derived labels and history-state targets are used only for training supervision, evaluation and audit. They are not model inputs at inference. Undocumented alternatives are treated as unknown rather than clinical negatives.

The output is a ranked candidate space grounded in the supplied clinical context. It is not a guideline engine, a calibrated patient-level risk model, a counterfactual treatment-effect estimator or an autonomous recommender.

Release contents

The release includes a 1.88-billion-parameter BF16 text backbone and the following separately packaged ClinTrace components:

the embedded text-only Qwen3.5-2B-Base language model with the compressor LoRA merged;
the frozen 32-token clinical-course compressor and renderer;
the compactness router and input plane;
the longitudinal history conditioner; and
the decision-navigation and conditional-outcome components.

manifest.json records the component layout, runtime contract and SHA-256 checksums. Although the repository root is loadable as a Transformers causal language model, that root represents only the merged compressor backbone. Running the complete ClinTrace pipeline requires the adjacent ClinTrace components and the iatro-clintrace runtime.

The release contains no clinical documents, patient identifiers, extracted chains, token caches, patient-level predictions or case-level evaluation outputs.

Development and evaluation data

ClinTrace was developed in a retrospective two-system cohort predominantly comprising HCC. The longitudinal modelling cohort contained 329,795 documents from 80,906 admissions in 9,725 patients. Source clinical records and patient-level derived assets are confidential and are not distributed.

Frozen evaluation

The primary completion endpoint was recall at five over observed P/X/O facets. Conditional-O metrics evaluated documented short-term feedback associated with each generated plan--action candidate.

Split	PXO R@5	O facet targets	O R@1	O R@3	O R@5	O MRR	O observed-set NLL
Validation	0.880236	3,551	0.615883	0.901436	0.957477	0.763672	1.141718
Internal test	0.879324	2,365	0.611416	0.897252	0.957294	0.759065	1.089216
External	0.808669	17,608	0.548046	0.877101	0.946047	0.719123	1.505285

In an independent review of 400 external cross-sections, five hepatobiliary oncology clinicians judged 76.5% of candidate pathways acceptable as presented (95% CI, 73.2--79.6). At least one acceptable candidate appeared within the top five in 92.9% of clinician--case evaluations (95% CI, 90.3--95.2).

These retrospective results measure recovery and clinical acceptability of candidate pathways. They do not establish effects on prospective decisions, workload or patient outcomes.

Limitations

The model was trained and evaluated under controlled data governance in two institutions within one national setting, and the cohort was predominantly HCC. Performance may shift across institutions, populations, diseases, documentation systems and clinical workflows. The model reflects patterns in documented care, including omissions and practice variation. Independent local validation, institutional governance and qualified clinician oversight are required before any clinical study or deployment.

Integrity

The selected final checkpoint SHA-256 is:

665fcd8779fa884101edee532060bd00bca35f659647c739bad6ed9b947ff709

Per-file release checksums are recorded in manifest.json.

License and attribution

The ClinTrace release artifacts are provided under CC BY 4.0. The embedded Qwen3.5-2B-Base originates from the Qwen Team and is distributed under the Apache License 2.0. Downstream use must preserve the applicable notices and attribution for both works.

Citation

@software{jia2026clintrace,
  author    = {Jia, Weili},
  title     = {iatrode/iatro-clintrace},
  version   = {1.0.3},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.21470033},
  url       = {https://doi.org/10.5281/zenodo.21470033}
}

Downloads last month: 2

Safetensors

Model size

2B params

Tensor type

BF16

Model tree for iatrode/iatro-clintrace

Base model

Qwen/Qwen3.5-2B-Base

Finetuned

(54)

this model