peal.adaptors

Adaptors: the procedures that repair a predictor from explanation feedback.

An adaptor takes a trained predictor and a source of feedback (a teacher) and returns a corrected predictor. This subpackage holds CFKD (counterfactual knowledge distillation on teacher-labelled counterfactuals), DiDAE (ranking of sparse-dictionary directions in a diffusion autoencoder’s latent space, followed by CFKD on the directions labelled false), the ClArC / P-ClArC / RR-ClArC concept projections, several Group-DRO variants and a projection adaptor. adaptor_factory.get_adaptor builds one from an AdaptorConfig yaml; every adaptor subclasses interfaces.Adaptor.

Submodules

adaptor_factory	Factory that resolves an adaptor specification into an Adaptor instance.
clarc	Class Artifact Compensation (ClArC) adaptors for removing known shortcuts.
counterfactual_knowledge_distillation	Counterfactual Knowledge Distillation (CFKD), PEAL's central repair loop.
didae	DiDAE: Dictionary-based Interpretable Diffusion Autoencoder Explanations.
group_distributionally_robust_optimization	Group Distributionally Robust Optimization (GroupDRO) as a PEAL adaptor.
group_dro_v2	Group Distributionally Robust Optimization as a PEAL repair adaptor.
interfaces	Base interfaces shared by every PEAL adaptor.
projection_adaptor	Adaptor that repairs a classifier by projecting concept directions out of its head.