CMI-RewardBench

dataset·music reward modeling·active

Preference data and evaluation for reward models scoring musicality, text alignment and compositional instruction following.

Verified facts

Official site	https://arxiv.org/abs/2603.00610 ↗
Geography	Global
arxiv	2603.00610
size	pseudo preference pairs: 110000
scope	Reward-model evaluation for text-conditioned music generation.
genres	multiple
steward	CMI-RewardBench authors
creators	CMI-RewardBench authors
languages	English prompts
modalities	music audio; text prompts; pairwise preferences; reward scores
geographies	Global
access method	Paper; linked data availability requires confirmation.
intended uses	music reward-model evaluation; preference modeling
consent claims	Human-rater consent and source-music permission details not verified.
related papers	CMI-RewardBench
annotation method	Preference judgments target musicality, text alignment and instruction following.
collection method	Combines large pseudo-preference data with a human preference set.
provenance claims	Paper describes CMI-Pref-Pseudo and human CMI-Pref components.
current availability	Paper available; durable corpus release not fully verified.
us market scope basis	Publicly available or materially used in US-facing music/audio-AI research.
limitations biases disputes	Pseudo-preference labels can reproduce evaluator-model bias.