CLEAR-MusicTags & ExpertEval-MusicTags

Music auto-tagging benchmarks such as MTG-Jamendo and MagnaTagATune carry weakly-supervised crowd or uploader tags. Their labels contain false positives, omit tags the audio does support, and treat near-duplicate tags inconsistently, which limits both training and evaluation.

This repository releases two artifacts, kept in distinct configs so that neither is implicitly tied to the other. They join on source_id.

• CLEAR-MusicTags (clear_*) — relabelled tags for the full train and test splits of five benchmarks, produced by a 2-of-4 consensus: a tag is kept when at least two of four annotation sources emit it.
• ExpertEval-MusicTags (expert_*) — a human-expert annotated evaluation subset, for measuring how well any label set agrees with expert judgement.

Both are described in DART: Difficulty-Aware Adaptive Reasoning for Music Information Retrieval (ACM MM 2026), which is the reference to cite.

No audio is distributed here. Only tags and stable clip identifiers are shipped; audio comes from MTG-Jamendo and MagnaTagATune directly, linked under Licensing and attribution.

Dataset statistics

Benchmark	Corpus	CLEAR train	CLEAR test	ExpertEval test	Tag vocab	Avg. tags/clip (original → CLEAR)
mtg_mood	MTG-Jamendo	9,949	4,231	1,000	56	1.79 → 6.82
mtg_instrument	MTG-Jamendo	14,395	5,115	999	39	2.44 → 6.32
mtg_genre	MTG-Jamendo	20,104	11,479	999	87	2.41 → 4.71
mtg_top50	MTG-Jamendo	20,006	11,356	999	49	3.07 → 10.58
mtt_top50	MagnaTagATune	18,706	5,329	400	36	2.66 → 7.82
Total		83,160	37,510	4,397

Train and test splits are disjoint at the clip level, and every ExpertEval clip is also present in the corresponding CLEAR test split.

mtg_top50 is a derived benchmark: its labels are the union of the mtg_mood, mtg_instrument and mtg_genre annotations for the same clip, restricted to the MTG-Jamendo Top-50 tag vocabulary. This holds for both the CLEAR and the ExpertEval labels.

Schema

Each row is one audio clip. Every column is a string.

Column	Configs	Description
id	all	Row identifier, <config>_<split>_<index>, unique across the whole release.
dataset	all	Benchmark this row belongs to: mtg_mood, mtg_instrument, mtg_genre, mtg_top50 or mtt_top50.
corpus	all	MTG (MTG-Jamendo) or MTT (MagnaTagATune).
source_id	all	MTG-Jamendo TRACK_ID (e.g. track_1028900), or the MagnaTagATune mp3_path. Joins the two artifacts to each other, and to the official corpus metadata.
audio_path	all	Path within the corpus audio archive: MTG-Jamendo audio-low, e.g. 00/1028900.low.mp3, or MagnaTagATune, e.g. f/clip.mp3.
tags_original	clear_*	Tags shipped with the original benchmark, with category prefixes such as mood/theme--- stripped. Empty where the original annotation is empty.
tags_clear	clear_*	CLEAR tags. Empty for 10 clips out of 120,670, where no tag reached consensus.
tags_expert	expert_*	Expert annotation. Never empty.

Tags are comma-separated, lower-cased and de-duplicated. Order carries no meaning — treat each field as a set. Each benchmark's closed vocabulary is also shipped as a plain list in tag_vocab/<benchmark>.txt.

Usage

A config name is required, since the repository holds ten of them.

from datasets import load_dataset

clear = load_dataset("lanw1/clear-musictags", "clear_mtg_mood")    # train + test
expert = load_dataset("lanw1/clear-musictags", "expert_mtg_mood")  # test only

Join the two on source_id to score either label set against the expert annotation.

Licensing and attribution

All tag annotations here are released under CC BY-NC-SA 4.0; use is limited to non-commercial research and academic purposes. The labels are modified from the sources below by consensus relabelling.

Configs	Source	Copyright	Annotation licence
*_mtg_*	The MTG-Jamendo Dataset	2019–2023 Music Technology Group, Universitat Pompeu Fabra	CC BY-NC-SA 4.0
*_mtt_top50	MagnaTagATune	Edith Law et al. (TagATune)	CC BY-NC-SA 3.0 Unported

No audio is redistributed; obtain it from the two sources above under their own per-track licences.