From Generality to Mastery: Composer-Style Conditioned Music Generation

Trained model weights and training datasets for the paper:

Mingyang Yao and Ke Chen "From Generality to Mastery: Composer-Style Symbolic Music Generation via Large-Scale Pre-training."
Conference of AI Music Creativity (AIMC), 2025

Note: Please find project details and usage at our Github repo

Model Architecture

"Generality" Stage

The model learns general music patterns and knowledge from diverse genres of music

Model backbone: 12-layer Transformer with relative positional encoding
Num trainable params: 39.6M

"Mastery" Stage

The model adapts its knowledge to specific composers' characteristics

Model backbone: 12-layer Transformer with relative positional encoding plus adapter modules inserted after every two transformer layers
Num trainable params: 46M

Citation

If you find this project useful, please cite our paper:

@inproceedings{generalitymastery2025,
  author = {Mingyang Yao and Ke Chen},
  title = {From Generality to Mastery: Composer-Style Symbolic Music Generation via Large-Scale Pre-training},
  booktitle={Proceedings of the AI Music Creativity, {AIMC}},
  year = {2025}
}