that low kl difference!!?
how did you achive 0 kl differnce, that's commendable, good work
The KLD is probably incorrect as per https://github.com/p-e-w/heretic/pull/160.
Also, this is among the initial batch of heretications prior to making refinements to this process (see latest releases). I think, there's still room to improve with this one, but at least waiting out the abovementioned PR is a logical step before re-processing. Thanks for the commendation.
but still, do you think using that technique it should be better than avarage abilitiration or heretic? with lower degradation
It's literally Heretic with MPOA, but with a manually optimised config to reduce degradation by applying the right amount of ablation weight, eliminating false-positive refusal detections, and adapting the markers to model-unique non-compliance and refusal patterns through studying its responses. This version was made with only the low-weight optimisation part of this. I cannot cast a judgement without proper evidence (I need feedback here people!), and plan on submitting some of these models that satisfied my vision to UGI evaluation to see if there's point to any of this.