MLP — UCI Heart Disease (Cleveland) — xaitalk tabular demo
A small 4-layer MLP (13→64→32→16→2) trained once on the UCI Heart Disease
(Cleveland) dataset, frozen and published so xaitalk's tabular adapter loads
fixed, reproducible weights (via xaitalk.hub.ensure_model) across
PyTorch / TensorFlow / JAX — rather than retraining on every load().
This makes the tabular domain deterministic (canonical + cross-device
consistency) and, crucially, demonstrates XAI on real, interpretable
clinical features (age, cholesterol, resting BP, ...) rather than synthetic
data — there is no signal to attribute in random data.
- Weights: framework-agnostic numpy
.npz (JAX layout; PT transposes Dense).
- Trainer: deterministic numpy Adam (seed 42), so the artifact is reproducible.
- Dataset: UCI Heart Disease (Cleveland), 303 samples, 13 features, 2 classes.
Part of
xaitalk — cross-framework XAI.
