CMI-Pref is a music preference comparison dataset for multimodal music generation research. Each record represents a single human vote comparing two generated audio samples, with preferences annotated along two dimensions (musicality and alignment) and a confidence score for each preference.
⚠️ Important Notes
The modality distribution below is computed over train + test combined.
The dataset contains overlapping votes by design: multiple users may vote… See the full description on the dataset page: https://huggingface.co/datasets/HaiwenXia/cmi-pref.
