You need to agree to share your Hugging Face information to access this model
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
Access is reviewed manually for research use. This model and its fitted tokenizer and probes derive from credentialed ICU databases. Describe your intended use and your authorization to use the relevant source datasets. Do not submit patient records, passwords, access tokens, or copies of restricted data in this form.
Log in or Sign Up to review the conditions and access this model content.
Didymos 1.0
Didymos is an autoregressive model of structured intensive care patient event streams.
This repository contains the frozen MUC32_main checkpoint described in
A generative digital twin forecasts and simulates intensive care patients without task labels.
Model
| Property | Value |
|---|---|
| Architecture | Qwen3ForCausalLM, trained on structured event tokens |
| Parameters | 123,491,328 |
| Layers / hidden width | 12 / 768 |
| Query / key-value heads | 12 / 4 |
| Feed-forward width | 3,072 |
| Vocabulary | 25,601 tokens |
| Training context | 32,768 tokens; the exported configuration supports 33,792 positions |
| Weights | float32, Safetensors |
| Training databases | MIMIC-IV, eICU-CRD, HiRID, AmsterdamUMCdb, SICdb, NWICU and INSPIRE |
The checkpoint was trained by next-token prediction without task labels. It is not a free-text chat model. Use the project's vocabulary, event grammar and sampling tools to prepare inputs and interpret outputs.
Access and permitted use
Access requests are reviewed individually. Access remains subject to the applicable source-dataset data-use agreements and the conditions of the authors' approval. This repository does not grant a general redistribution license for the weights, fitted tokenizer statistics or probes. The Apache-2.0 license of the source-code repository applies to that code, not automatically to these model artifacts.
The model is for research, not clinical decision-making. Records, token shards, patient-level outputs and synthetic stays are not included. The checkpoint adapted to RICD and zigongICU is also not included.
Files
| Path | Contents |
|---|---|
hf/ |
Model configuration, generation settings and model.safetensors |
tokenizer/ |
Custom vocabulary, value bins, decoding statistics and dataset support |
audit/ |
Supplied privacy-audit reports |
toolkit/recalibration/ |
Recalibration maps used by the toolkit |
paper/a1/test_{L24,L48,R}/ |
Probe coefficients used for model-state explanations |
FROZEN.json |
Original training configuration and freeze record |
train_config.json, train_report.json, metrics.jsonl |
Training settings and records |
Paths inside the historical training records describe the training host. They are provenance records, not required local installation paths.
Load after approval
Use Python 3.12 and install the source repository's locked environment with
uv sync --frozen. Authenticate with your own Hugging Face account after approval.
from pathlib import Path
from huggingface_hub import snapshot_download
from transformers import AutoModelForCausalLM
from icutwin.generate.producer import require_gate
from icutwin.tokenization.vocab import Vocab
root = Path(snapshot_download("HanZiwww/Didymos-1.0"))
require_gate(root / "audit", root / "hf")
vocab = Vocab.load(root / "tokenizer" / "vocab.json")
model = AutoModelForCausalLM.from_pretrained(root / "hf").eval()
The tokenizer is the project's Vocab, not a standard free-text AutoTokenizer.
For forecasting and constrained simulations, follow
the toolkit documentation
and supply locally authorized, harmonized MEDS inputs.
Evaluation limits
The supplied audit records a PASS gate for these exact weights. Its prompted
sampling assessment also records an inconclusive, sample-limited INSPIRE result;
the detailed report and confidence intervals remain in audit/. These records do
not establish an absence of privacy risk or override source-data agreements.
The manuscript reports that unrestricted generation underproduces or postpones deaths. Forecasting, free-running trajectories and treatment-response comparisons have different validation limits; consult the source repository's methods and results before reusing them.
Provenance
The weight file was preserved from didymos_project_2026-10-14.zip.
Its SHA-256 is
f8fe56d6d7c65abea9353cc9d0494077d5b404677d344da4cc55fa20d7253098.
The release source snapshot is recorded in the code repository's COMMIT.txt;
FROZEN.json records the earlier development commit used during training.