For more information on Editing, see point 3 in #112 .
Called ablation, but perform masking of features using a baseline.
Editing replaces tokens with their nearest neighbors in the vocabulary embedding space and measures saliency as the drop in performance for the target. In the future, this can allow users to specify a custom editing strategy via an input Callable.
Possibly overlapping with feature ablation up to some measure.
Valid only for decoder-only models.
Verify whether it would be exactly equivalent to Value Zeroing, include only if functionally different (alias otherwise).
🚀 Feature Request
The following is a non-exhaustive list of perturbation-based feature attribution methods that could be added to the library:
pytorch/captum
pytorch/captum
pytorch/captum
pytorch/captum
pytorch/captum
keyonvafa/sequential-rationales
DFKI-NLP/thermostat
dylan-slack/Modeling-Uncertainty-Local-Explainability
dylan-slack/Modeling-Uncertainty-Local-Explainability
DFKI-NLP/OLM
cifkao/context-probing
ykwon0407/WeightedSHAP
hmohebbi/ValueZeroing
YilunZhou/solvability-explainer
YilunZhou/solvability-explainer
kmeng01/rome
casszhao/ReAGent
k-amara/syntax-shap
Notes:
Callable
.