CMI-Pref: Human Music Preference Comparisons with Alignment Scores

CMI-Pref provides between 1,000 and 10,000 human preference comparisons for multimodal music generation, published by HaiwenXia in 2026. Each record captures a human vote comparing two generated audio samples based on musicality, alignment, and confidence.

Use Cases

• Training reward models for RLHF using the musicality and alignment preference dimensions
• Evaluating audio generation quality by weighting samples with the provided confidence scores
• Analyzing annotator consistency using the overlapping vote design where multiple users evaluate the same pair

Strengths

• Provides dual-axis labels for musicality and alignment rather than a single preference score
• Includes confidence scores for every human judgment record
• Features an overlapping vote design to support inter-annotator reliability analysis

Limitations

• Limited scale of 1,000 to 10,000 records
• Restricted to non-commercial applications under the CC BY-NC-SA 4.0 license

Provenance

Source

HaiwenXia

Collection Method

Human annotation

Freshness

Last updated March 2026