Fit LGP Configuration

bff fit-lgp trains one local Gaussian-process committee for each QoI dataset.

datasets:
  rdf:
    data: ../04-qoi/qoi/rdf.pt
    mean: sigmoid
fit:
  model_dir: ./models
  reuse_models: true
  n_hyper_max: 200
  committee_size: 1
  test_fraction: 0.2
  device: cuda

datasets.<name>.data must contain a dataset with the same name. Optional dataset keys are mean, nuisance, and model; the model path defaults to fit.model_dir/<name>.lgp. Optimizer overrides are lr, max_iter, and tol_grad. Unknown keys are rejected.

The default log is fit-lgp.log next to models/. Cached models include an exact fingerprint of the QoI dataset. Reuse is rejected if any training input, output, reference value, label, setting, or metadata differs, even when tensor shapes are unchanged.