DiVers: A large and diverse version identification dataset.

This collection of version identification dataset is based on the cliques in Discogs-VI (DVI) and contains more than 700k additional versions from YouTube. We present the YVI datasets, which contain newly added versions and DiVers datasets, which are the union of YVI and DVI. Datasets small DiVers: > 700k versions YVI: > 200k versions large DiVers > 1.3M versions YVI > 900k versions Structure Keys: first-level keys are the subsets (train, valid, test) second-level keys are the clique identifiers third-level keys are the version identifiers Subdictionaries per version identifier contain for the light files (will be sufficient for most use cases): id: integer version identifier artist: release artist (only for DVI versions) title: release title (only for DVI versions) filename: filename of the MP3 youtube_id: YouTube identifier dvi: indicates if version is a DVI version (from DiscogsVI) tags_yt_title: music tags detected in the YouTube video title the other files contain rich metadata (caution! some of these are often None) samplerate: sample rate length: duration in seconds channels: number of channels track_writer_names: list of writers of the track (only for DVI versions) release_artist_names: list of artists of the track (only for DVI versions) release_genres: list of genres (only for DVI versions) release_styles: list of sub-genres (only for DVI versions) country: release country or region (only for DVI versions) labels: music labels formats: release formats (e.g., CD, Vinyl) released: release year yt_title: YouTube video title yt_tags: YouTube tags yt_categories: YouTube categories yt_channel: YouTube channel cues_yt_title: cues for music tag detection in the YouTube video title tags_yt_description: music tags detected in the YouTube description cues_yt_description: cues for music tag detection in the YouTube video title tags_yt_tags: music tags detected in the YouTube tags cues_yt_tags: cues for music tag detection in the YouTube tags