gerstung-lab/Delphi — reverse-engineered prompt

Build me a research prototype for modelling human disease histories with a small generative transformer, inspired by Delphi. I want to be able to train it on a prepared patient event dataset, and also include a tiny synthetic sample dataset so the whole thing can run without private medical data. Please make the setup easy with a Python environment, requirements file, simple training script, and clear configs for a demo run on GPU or CPU.

After training, I want notebooks that let me evaluate prediction accuracy against a simple age and sex baseline, check calibration, look at attention patterns and disease embeddings, run SHAP style explanations for patient history importance, and sample or compare generated synthetic health trajectories. Include a helper example for turning UK Biobank style raw records or similar custom data into the format the model expects. Add a README that explains installation, training, Docker use if practical, data limitations, and how to cite the related paper. Look up current docs online if needed.