Pop-K is a generative MIDI model that creates pop melodies in C major and A minor. Built on RWKV-4, the architecture employs a pseudo-linear attention mechanism for efficient training and fast inference. The model was trained on ~300k 8-bar melody samples composed of augmented bass, chords and vocal/lead melodies.
Download the checkpoint from Hugging Face:
Use colab_run.ipynb to run the model with Gradio or generate MIDI files in bulk.
Preview MIDI outputs in data/midi_output_examples.zip.
The Pop-K MIDI Dataset is closed-source and not publicly available for direct download. However, more information about the dataset, including details on access and usage, can be found at the following link:
