SocialNLI is a dialogue-centric natural language inference benchmark that probes whether models can detect sarcasm, irony, unstated intentions, and other subtle types of social reasoning. Every record pairs a multi-party transcript from the television series Friends with a free-form hypothesis and counterfactual explanations that argue for and against the hypothesis.
