Peter shares how they created a different pretraining set of data by scraping case law data and turning it into a multiple choice task for holding statements. He explains the importance of contextual cues in crafting persuasive holding statements and how pretraining on this data helps models pick up on these cues.
