Fully automatic censorship removal for language models
Commits on master
Perform calculations involving residual vectors in float32
a9655c8 Philipp Emanuel Weidmann committed 1 year ago
Dynamically choose between global and per-layer refusal directions
1496e0a Philipp Emanuel Weidmann committed 1 year ago
Add functionality to evaluate any model relative to the main model
cf57a0c Philipp Emanuel Weidmann committed 1 year ago
Separate abliteration parameters for different layer components
6359aa4 Philipp Emanuel Weidmann committed 1 year ago
Support passing model name without "--model" argument prefix
7573a2e Philipp Emanuel Weidmann committed 1 year ago