merge disparate task vectors
Are you planning on merging all of these at the end? It's quite an interesting little project you have here.
Likely not, they're being developed to try to have model organisms with diverse values. Still needs at least one more pass on top of these with some additional data to start to show those reliably, I think.
Since the axes are specifically trained against each others' polarities, task vectors would likely conflict/cancel out to a fair degree.
That's what makes it so interesting. Orthogonal datasets create the most distinct task vectors and have the most powerful effects when merged; particularly if you are doing SFT then DPO/GRPO/GSPO (or whatever the policy optimization flavor of the moment is).
Opposed/reversed is not quite orthogonal, though it could certainly be interesting to see the results when merging.
Luminous-Designs/Qwen3.5-9B-Schwartz-TIES is interestingly different from Thoughtful-Interiority; do seem balanced and show flashes of all the value poles, as well as heightened emotional register. Orthogonal enough.