peal.sparse_dictionaries.probe_sparse_autoencoder¶
Sparse dictionary made of the weight rows of a pretrained linear probe.
ProbeSAE is the degenerate member of PEAL’s sparse dictionary family: instead
of learning atoms from activations it takes the final fc layer of a trained
classifier checkpoint and exposes its class weight vectors as the components. It
lets the DiDAE / CFKD pipelines use the classifier’s own decision directions as
candidate edit directions without any fitting step.
Classes
Dictionary whose atoms are the rows of a classifier's last linear layer. |
|
Config of |