chackochigreat/semscore-notagen-ft-pdmx

semscore piano model (NotaGen fine-tune)

These are the music-writing model weights for semscore, a system that turns any English or German text into an engraved piano score. Built for the seminar "Semantic-Driven Music Score Generation" (UMTL, Saarland University / DFKI, Summer 2026).

The base model is NotaGen, a 244-million-parameter transformer that writes music in ABC notation. We fine-tuned it on piano pieces from the public-domain part of the PDMX score collection. During fine-tuning, every training excerpt was labelled with the musical constraints semscore works with (which notes are allowed, how high or low, how many notes per bar, what texture). The measured gain is on the note set: on raw output, the share of bars that stay inside the requested note set rises from 28% (base model) to 66% (this fine-tune). The other constraint kinds did not improve the same way; in semscore a separate checking layer guarantees them.

Files

File	What it is
ft2a_attr_aug_e3.pth	The model semscore uses. Trained with constraint labels and transposition augmentation.
ft2b_attr_e2.pth	Comparison model: same training but without the augmentation. Kept so the effect of augmentation can be checked (66% vs 57% on the note-set measure).
ft_pdmx_e1.pth	First fine-tuning attempt, without constraint labels. Kept for comparison; it adapts to the piano style but does not learn to follow constraints.

SHA-256 checksums:

• ft2a_attr_aug_e3.pth: 88e664155f884ee7872a3a4da004b365f516c787a6da4fad058897c54aa6692a
• ft2b_attr_e2.pth: 101749a48a006cf7606bb0cb880ee80cb86a46062908f2695180bab7714a7a93
• ft_pdmx_e1.pth: ad9fd66d120b9e6652bb1770fd096c013affb1441a9ef9c4050e05ce5a51eb29

How to use

Run python scripts/setup_weights.py in the semscore repository. It downloads the base model (from ElectricAlexis/NotaGen) and ft2a_attr_aug_e3.pth (from this repository), verifies the checksums, and puts everything where semscore expects it. Nothing needs to be done by hand.

Licences

• Base model weights and inference code: MIT (ElectricAlexis/NotaGen). The base model's own pretraining data is not fully disclosed by its authors.
• Fine-tuning data: public-domain subset of the PDMX score collection.