Multi-Sarcasm is a 23,715-sample conversational dataset of parent–reply pairs, combining Reddit and chatbot dialogues with graded sarcasm annotations.