The Topology and Orchestration Specification for Cloud Applications (TOSCA) is the OASIS standard for portable cloud orchestration, yet a decade of adoption has produced a fragmented landscape of mutually incompatible dialects and, remarkably, no public dataset with which to study any of it. We address this with a curated dataset of 14,931 real-world TOSCA blueprints drawn from 260 open-source repositories, identified among 1,973 candidates through three complementary discovery strategies (repository keyword, file-content, and topic search). Because no single parser spans every dialect, we curate the corpus with a three-tier, parser-independent quality model, i.e. parseability, structural coherence, and version classification, and annotate each file with provenance, dialect and version labels, and structural-complexity metrics. The result is a multi-dialect snapshot of TOSCA as practitioners actually write it: TOSCA 2.0 dominates (54.1%), but Simple Profile 1.x (33.6%) and engine-specific DSLs such as Cloudify and Alien4Cloud endure, alongside a long tail of vendor and academic extensions missing from the official specifications. The dataset supports empirical studies of TOSCA evolution, cross-project reuse, structural quality, and version migration. Concretely, it can also serve as training and evaluation data for LLM-based blueprint generation, repair, and version 1.x to 2.0 migration, and as ground truth for dialect classification, clone detection, and cross-dialect parser benchmarking.