DiVers-1M: A large and diverse version identification dataset.

This version identification dataset is based on Discogs-VI and contains more than 700k additional versions from YouTube. In this repository, we provide three datasets with compatible splits: dvi2: a second, cleaned version of Discogs-VI yvi: a YouTube crawl comprising newly found versions on YouTube divers1m: the combination of both datasets File Structure we include a subdir with Torch and one with JSON files with the same data each dataset is a dictionary with the following content: info: global version IDs mapping to the version metadata global ID is the clique ID concatenated with the version ID with a ":" in between. Example: C-01:V-07 with clique ID C-01 and version ID V0-7 depending on the type of file (indicated by the sub-directory), we either have light: only basic metadata on the song level and the sampling rate and length rich: everything in light but also the matched concepts, YouTube title and tags, etc. split: split key (train, valid, test) mapping to the clique identifiers which map to lists of version identifiers