Even if they did, I’d assume the association of “full” and this correct representation would benefit other areas of the model. I.e., there could (/should?) be g

Even if they did, I’d assume the association of “full” and this correct representation would benefit other areas of the model. I.e., there could (/should?) be general improvement for prompts where objects have unusual adjectives. So maybe training for litmus tests isn’t the worst strategy in the absence of another entire internet of training data…