anyone who wants to train an AI model on relatively authentic data and natural (for certain definitions of "normal") human interactions, I would assume like ID

anyone who wants to train an AI model on relatively authentic data and natural (for certain definitions of "normal") human interactions, I would assume like IDK what the comps are, what is another large corpus with authentic interactions and scoring? internet archive is probably the closest (although substantially more valuable), maybe the google library project is semi analogous in scale and value (although obviously very different in nature).