Discogs-VI-SIREN: SImilar REgioNs between musical versions in Discogs-VI

Discogs-VI-SIREN contains SImilar REgioNs between musical versions in Discogs-VI (https://zenodo.org/records/13983028). For every pair of versions of the same work, the audio is embedded in 20-second segments at a 1-second hop using CLEWS embeddings, a pairwise distance matrix is computed between the two versions, and the local minima of that matrix are grown into basins: regions where the two recordings stay close to each other. similar-regions.csv lists one basin per row, together with the best-matching pair of 20-second segments and its distance (96,754,487 rows for train, 9,564,538 for val). segment-cliques.csv groups the resulting segments across all versions of a work into segment-level cliques, i.e. sets of segments from different versions that render the same passage of music (18,566,717 and 2,099,049 rows). The train and val splits follow those of Discogs-VI-YT and share no works. Each split also ships the parameters used to generate it. The audio itself is not included; obtain it through Discogs-VI-YT. Released under the MIT License. Code: https://github.com/raraz15/unified-track-and-version-id Please cite the accompanying ISMIR 2026 paper, "Unified Music Identification for Tracks and Versions" (Araz, Serrà, Mitsufuji, Serra, Bogdanov), and the CLEWS paper it builds on (Serrà, Araz, Bogdanov, Mitsufuji, ICML 2025). Full details are in the README inside the archive.