Retrieval shortlists for training-free k-reciprocal re-ranking in musical version identification

Top-K retrieval shortlists and per-query records for Training-free k-reciprocal re-ranking improves state-of-the-art musical version identification. Each dump holds, for every track of a Discogs-VI split used as a query, the ranked ids of its nearest neighbours, their distances, and whether each candidate is a true version. The per-query records hold average precision before and after re-ranking, the rank of the first correct hit, and the clique size. No audio and no embeddings are included, and none can be reconstructed from these files. The files are derived from the Discogs-VI dataset, whose metadata is licensed CC BY-NC-SA 4.0, and are released under the same licence. Please cite R. O. Araz, X. Serra and D. Bogdanov, Discogs-VI: A musical version identification dataset based on public editorial metadata, ISMIR 2024, alongside this record. Code: github.com/pyroxenite/k-reciprocal-vi.