peal.sparse_dictionaries.probe_sparse_autoencoder.ProbeSAEConfig

class peal.sparse_dictionaries.probe_sparse_autoencoder.ProbeSAEConfig[source]

Bases: SparseDictionaryConfig

Config of ProbeSAE.

Parameters:
  • pretrained_model_path (str) – Path of a torch.save-d Lightning predictor whose model.model.fc is the linear layer whose weight rows become the dictionary atoms.

  • n_components (int or None, default 2) – Number of components; informational only, the probe decides the count.

  • sparse_dictionaries_type (str) – Registry key, fixed to "ProbeSAE".

  • ending (str) – File ending the factory expects for saved dictionaries (".npz").

Methods

Attributes

pretrained_model_path: str
n_components: int | None = 2
sparse_dictionaries_type: str = 'ProbeSAE'
ending: str = '.npz'