Examples#
The repository has runnable demonstrations that mirror typical HEP optimization tasks.
Each script generates toy datasets, trains a gato-hep model, and writes diagnostic plots/tables under the corresponding examples/.../Plots*/ directory.
Specifically:#
1D sigmoid/GMM - compare different approaches to sculpting categories in a single discriminant and inspect penalty terms, bias, and yield-vs-uncertainty plots.
Three-class softmax - operate directly on a three-class score. Can be similarly used on multiple 1D discriminants by stacking their scores.
Diphoton bump-hunt - uses the three-class softmax problem, here mimicing a Higgs-to-γγ workflow: add a diphoton-mass observable, fit continuum sidebands with exponentials to use the full continuum bkg. statistics but reweight to the signal window fraction.