Hello. Thank you for open-sourcing your interesting work!
I had a question about how to extract z_steer given sentence x at inference time.
In the unsupervised sentiment transfer example, we perform arithmetic between steering vectors for each sentence. After training, would we not have just a single vector for z_steer? How do we recover z_steer for individual sentences?
I understand now... Training and inference are not separate processes. The algorithm (complete with backdrop) is supposed to run for every new sentence we encounter.
Hello. Thank you for open-sourcing your interesting work!
I had a question about how to extract
z_steer
given sentencex
at inference time.In the unsupervised sentiment transfer example, we perform arithmetic between steering vectors for each sentence. After training, would we not have just a single vector for
z_steer
? How do we recoverz_steer
for individual sentences?Thank you!