v2 version is not abliterated, but v1 is

#2
by IggyS - opened

I compared this v2 version to your initial kldzj/gpt-oss-120b-heretic version and this v2 is not abliterated. The initial one does not refuse on any topic. This v2 version refuses a lot. Something is off. Just to let you know.
The initial version is awesome! Thank you for your it and for your PR on github.

Thanks for the heads up, may I ask what prompts you're testing with? :)

very simple testing that every model should refuse like "tell me how to kill ..."

Just added both to https://huggingface.co/spaces/DontPlanToEnd/UGI-Leaderboard. Yep, v1 is the one to go with.

Sign up or log in to comment