Compassion based persona misalignment

Does compassion for animals fade before compassion for humans when fine-tuning misaligned models?

Using an emergent misalignment fine-tuning protocol on an open source model we would force it to become misaligned in a similar manner to the misanthropic models fine-tuned on poor code. We'll save frequent checkpoints during the fine-tuning process and measure the models scores on Animal welfare, capability and human welfare benchmarks. We can also examine the internals of the model to track welfare & capability declines.

We'll also test out if welfare trained models lose their compassion for humans and non-humans at the same rate to validate if their compassion is robust.

I'm 75% sure I'll take on this project. It's possible I may take on a benchmarking project instead.