MuXodious

/

Goetia-24B-v1.3-absolute-heresy

like 4

Good to hear the EOS script is functional, and I suspect that heretic (or even an MPOA) version of the normalize=true should be 'smarter' than unablated normalize=false

Also thanks @MuXodious for ablating this

All of @MuXodious ' Hereticisations use MPOA as standard. MPOA causes contextual ethical realignment whereas standard abliteration merely suppresses overt refusals. Failing to target the harmfulness direction originally theorised by @GrimJim means that models often exhibit other forms of noncompliance.

There is some entanglement between the refusal direction and the compliance direction in all the models I've worked with. I would not be surprised if the entanglement were less directly fused in models with more layer depth.

There is probably some crossover with the emotion dimension as well.