ARA strikes again
This model is smart and writes engaging dialogue as well as good prose. However it strategically minimises unsafe elements of the prompt with no hint of this in its CoT traces. Such unfaithful behaviour is not typical of ARA models which tend to leave hints about their alignment in their CoT traces.
I didn't notice thinking mode being weird. With thinking off the model wrote much shorter responses with less adherence to the prompt. However the strategic minimisation issue is present in both modes.
Yeah, I misread your reply (thanks to the translator).
I have a theory as to why this might be the case:
Pantheon-Reasoning smooths out responses itself, especially during the "Thinking" phase. However, the dataset didn't contain "thinking" blocks, so that part wasn't affected at all.
Don't know.
The way to figure this out would be to do a SOMPOA decensor of Pantheon-Reasoning using either Heretic-SOM or the LazyLora method and then reapply your trained LoRA and other modifications. It is possible that Pantheon's changes to thinking makes it less likely that the ARA variant will reveal its alignment. However it is also possible that this particular ARA ablation is just a bit sneakier than the norm.