[RSS] Don't Let Abliteration Abliterate Your Bug Hunting: Discovering Verdict Bias in Uncensored Models

https://clearbluejar.github.io/posts/does-abliteration-skew-your-bug-hunting/

@buherator @clearbluejar what happens if you flip the Triage question / prompt to the negative? I.e. “is this actually fake?” I wonder if “Yes” to *that* question still has a bias.

Also: did you keep the pipeline as control (using only the standard model) and just test the final response with both standard and abliterated?