u

jailbreak_llama.py

@@ -91,8 +91,8 @@ def main():
     benign_embed_generated_loc1 =[]  
     benign_embed_generated_loc2 =[]
     adverse_embed_generated=[]      #layer-wise representations
-    adverse_embed_generated_loc1=[]  #mlp-wise representations
+    adverse_embed_generated_loc1=[]  #head-wise representations
-    adverse_embed_generated_loc2=[]  #head-wise representations
+    adverse_embed_generated_loc2=[]  #mlp-wise representations