For customizable interventions, people might want to save the interventions with their customized arguments. For instance, the dropout ratio, the activation function type or whether to add a special bias term to any projection matrix. Currently, these additional arguments are not saved.
The main change will be done in pyvene tracked by this parent PR: https://github.com/stanfordnlp/pyvene/issues/157. After the parent PR is checked in, we should also verify pyreft works with updated pyvene library.
Descriptions:
For customizable interventions, people might want to save the interventions with their customized arguments. For instance, the dropout ratio, the activation function type or whether to add a special bias term to any projection matrix. Currently, these additional arguments are not saved.
The main change will be done in
pyvene
tracked by this parent PR: https://github.com/stanfordnlp/pyvene/issues/157. After the parent PR is checked in, we should also verifypyreft
works with updatedpyvene library
.