Defensible sets and stored partitions for "Multiverse analysis of embedding-based text clustering and label-free selection by agreement"

This dataset accompanies the article "Multiverse analysis of embedding-based text clustering and label-free selection by agreement". It holds the two largest outputs of the study, which are too large to be supplementary files of the article. Raw_data_defensible_sets.zip (138 MB) holds the defensible set of each of the 63 corpus samples (21 corpora × 3 samples), one .npz file per sample. Each file contains: - the defensible specifications; - their pairwise ARI and AMI agreement matrices; - the ARI of each member against the gold labels; - the MAS medoid and the MAS consensus partition; - the median K; - the cross-space silhouette scores; - the planted-shift recovery of the members, the consensus and the gold labels.Raw_data_partitions.zip (1.2 GB) holds every stored partition: the cluster label vector and scores of every fitted clustering, with one .npz file per corpus sample, representation, reduction and seed. It covers the main multiverse (runs/), the analysis with samples of 1000 documents (runs_n1000/) and the extended multiverse (runs_ext/), together with the defensible sets of the last two analyses.No corpus text is included. The corpora are public and are downloaded from the Hugging Face Hub by the code supplied with the article, and the samples are drawn with fixed seeds, so every document set can be rebuilt exactly. README_DATA.txt in each archive describes every file and field. The per-specification scores, result tables and code are supplementary files of the article.