MARBLE

dataset·music audio representation·active

A unified evaluation suite for pretrained music-audio representations across acoustic, performance, score and high-level tasks.

Verified facts

Official site	https://github.com/a43992899/MARBLE-Benchmark ↗
Geography	Global
arxiv	2306.10548
size	v2 tasks: 18 · v2 datasets: 12 · v2 task categories: 13
scope	Standardized probing and downstream evaluation of music-audio representations.
genres	multiple public-source corpora
license	Code license plus source-dataset-specific terms
steward	MARBLE research collaboration
creators	Ruibin Yuan; Yinghao Ma; Gus Xia; Jie Fu
versions	v1; v2
languages	mostly not applicable; lyric tasks are language-dependent
modalities	audio; task labels; evaluation code; model checkpoints
geographies	Global
access method	Public GitHub repository; some dataset audio must be obtained separately.
intended uses	music representation evaluation; MIR benchmarking
consent claims	Inherited from source datasets; not uniform.
related papers	MARBLE: Music Audio Representation Benchmark for Universal Evaluation
prohibited uses	Any use prohibited by incorporated source licenses.
annotation method	Reuses source-dataset labels and standardized train, validation and test protocols.
collection method	Integrates public music datasets under a common configuration and evaluation protocol.
provenance claims	Repository documents each incorporated dataset and notes copyright constraints.
current availability	MARBLE v2 is on the main branch.
commercial use limits	Commercial suitability varies by incorporated dataset.