Adds a new method argument to ControlVector.train. This defaults to the old method, method="pca_diff", but it can be switched to method="pca_center" for more stability. This seems to be necessary for training vectors on larger models, such as llama3-70b. Thanks to @r3ndd
Adds ControlVector.import_gguf as a counterpart to ControlVector.export_gguf. GGUF is now the preferred way to save control vectors, for better compatibility with llama.cpp (however, the old pickle method will continue to work)
Changes
method
argument toControlVector.train
. This defaults to the old method,method="pca_diff"
, but it can be switched tomethod="pca_center"
for more stability. This seems to be necessary for training vectors on larger models, such as llama3-70b. Thanks to @r3nddControlVector.import_gguf
as a counterpart toControlVector.export_gguf
. GGUF is now the preferred way to save control vectors, for better compatibility with llama.cpp (however, the old pickle method will continue to work)ControlVector
Co-Authored-With: r3ndd@users.noreply.github.com