in the info prompt without inaccurate location case, the logit lens' predictions between layers 60 and 63 have nearly *perfect* accuracy - ~100% yes if there was an injection ~0% if there wasn't. that's pretty remarkable.
in reply to: 2002627632626819486
same thread: 2002601461969047902 2002627632626819486 2002631862335013094 2002636727291425196 2002644492776845371 2002648130396041550
Reproduced against link rot, credited and linked to its original. Yours and you’d rather it weren’t here? Open an issue.