I was the Dataset and Annotation lead. I selected three public-domain US speech excerpts (Gettysburg Address, Second Inaugural, FDR First Inaugural) and built the contrastive dataset: for each text, 21 items covering near-perfect mirrors, six single-factor flaw families (fast and slow pace, excess and missing pauses, flat pitch, quiet energy) at three severities, and a three-flaw combination. I defined the dataset schema, stable item IDs, provenance fields and exact ground-truth flaw spans, and documented the dataset and its limits. I kept the held-out speeches separate from threshold calibration so the evaluation stays honest.
Log in or sign up for Devpost to join the conversation.