CoPaGE-300K: Controlled Paraphrase Geometry Embeddings

CoPaGE-300K (Controlled Paraphrase Geometry Embeddings) is a controlled slot-based dataset designed for studying the local geometry of semantically close sentence embeddings. The dataset contains more than 300,000 template-generated English sentences organized into three template families and five slot-variation regimes. Each sentence is associated with explicit slot metadata and precomputed sentence embeddings. The dataset is intended for experiments on local manifold-like structure, affine, quadratic and cubic fitting of local embedding clouds, manifold-aware latent augmentation, surface-based generation of synthetic latent points, and the comparison between geometric validity and downstream discriminative utility. CoPaGE-300K should not be interpreted as a natural paraphrase corpus. Its value lies in its controlled factorial structure, explicit slot metadata, reproducible generation procedure, and suitability for local geometric analysis in sentence embedding spaces.