Which pLM to choose? Embeddings, protein pairs and analysis data

Data for the manuscript Which pLM to choose? (preprint: doi:10.1101/2025.10.30.685515; code: github.com/tsenoner/plm_choice). The study compares 14 protein language models (pLMs) and a random Gaussian control on 75,849,972 Swiss-Prot protein pairs, each scored for sequence identity (PIDE), HFSP and structural similarity (TM-score). It separates the signal already in raw embedding distances from the signal a supervised probe extracts, and compares embeddings with an MMseqs2 homology search for transferring EC numbers and GO molecular-function terms. This record holds: per-protein embeddings of 540,881 Swiss-Prot 2024_01 proteins for all 15 embedding sets (embeddings_*.h5); the deduplicated pair set with its training, validation and test splits, pair counts and an identical-sequence mask; the probe grids (trained probes and raw-distance scores for 3 targets) on native embeddings, 128-dimensional PCA projections and 11 untrained networks; the EC and GO-MF cohorts with their annotation-transfer results and the MMseqs2 homology baseline; test pairs rescored from experimental PDB structures; per-model distance histograms, and 5,000,000 random pairs with their distances under every model; the New2024 test set of proteins new in UniRef50 2025_01. Start with README.md; every file format is in DATA_DICTIONARY.md. Version 2 accompanies the revised manuscript. It updates the embeddings, pair tables and probes and adds the analysis data above. Version 1 (2025) remains available. Cite all versions with the concept DOI 10.5281/zenodo.17469267.