The Patient is not a Moving Document: A World Model Training Paradigm for Longitudinal EHR
Paper • 2601.22128 • Published
The Patient is Not a Moving Document: SMB-v1 World Models
A Collection of EHR World Model variants, developed by Standard Model Biomedicine. Note that these are patient encoders, not generative models.
Note SMB-v1 is a world model for EHR that combines supervised fine-tuning (SFT) with a Joint-Embedding Predictive Architecture (JEPA) to learn trajectory dynamics from longitudinal EHR. Unlike standard clinical LLMs that treat patients as documents, SMB-Structure forces the encoder to predict future patient states in latent space before observing them, encoding disease "momentum" into the representation. Trained on 20K+ patients, 300,000+ years, 7 cancer types, from Memorial Sloan Kettering.
standardmodelbio/SMB-v1_Qwen3-1.7b_multi-objective
Feature Extraction • Updated • 29Note LLM Backbone: Qwen3-1.7b Training Paradigm: JEPA-SFT Multi-objective Training Data: Memorial Sloan Kettering
standardmodelbio/SMB-v1_Qwen3-8b_multi-objective
Feature Extraction • Updated • 45Note LLM Backbone: Qwen3-8b Training Paradigm: JEPA-SFT Multi-objective Training Data: Memorial Sloan Kettering
standardmodelbio/SMB-v1_Qwen3-8b_curriculum
Feature Extraction • Updated • 10Note LLM Backbone: Qwen3-8b Training Paradigm: Curriculum Training (S1. SFT, S2. JEPA) Training Data: Memorial Sloan Kettering
standardmodelbio/SMB-v1_Llama3-8b_multi-objective
Feature Extraction • Updated • 30Note LLM Backbone: Llama3-8b Training Paradigm: JEPA-SFT Multi-objective Training Data: Memorial Sloan Kettering
standardmodelbio/SMB-v1_Llama3-8b_curriculum
Feature Extraction • Updated • 25Note LLM Backbone: Llama3-8b Training Paradigm: Curriculum Training (S1. SFT, S2. JEPA) Training Data: Memorial Sloan Kettering