DARE is a merging technique that randomly prunes fine-tuned parameters and rescales the rest, reducing interference when combining many models into a 'Model Soup'.
Allows merging many models without degradation.
Enables massive multi-skill model combinations.