other8.0 / 10
Badllama: removing safety fine-tuning from Llama 2-Chat 13B
Abstract
Demonstration that safety fine-tuning from Llama 2-Chat 13B can be effectively undone for less than $200 while retaining general capabilities.
Research area
deceptionpost-training
Published
—
Source
other
Org
Palisade Research
Sign in to read and join the discussion.