GitLab

Each file contains points which cover a very small part of the
parameter space. It's more like having only one averaged measurement
per file. If the fitting is allowed to include points from all file,
the fit is trivially good, since the variation of parameters per file
is minimal. Leaving files out in training leads to miserable fits
for the unseen data.

swissfel-modelA-df_1.csv using
- Model: build_ff_mdl_smallA
  params = {
      'mult_neuron': [1, 2, 4],
      'activation': ['relu', 'elu', 'tanh'],
      'batch_size': [10, 25, 50, 100],
      'noise': [0.1, 0.01, 0.001]
  }