So... distribute a LoRA (or equivalent) that modifies the base weights with the abliteration vectors. That makes sense as it would be possible to try different abliterations and keep the storage space down.

▲

wren6991 5 hours ago | parent [-]

Yeah, orthogonalisation is a special case of a rank-1 LoRA.

▲

ziofill 2 hours ago | parent [-]

I don’t think this is true.

	▲	wren6991 an hour ago \| parent [-]
		Could you please explain why? It's an additive update to the weights, adding an outer product of a vector with its transpose, which must have rank 1. What am I missing?