Comment by rhdunn

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.