Music Transformer (ATEPP pretrain)
Unconditioned Music Transformer trained on ATEPP-1.2 piano MIDI. This is the
default evaluation checkpoint atepp-pretrain in
expectation-benchmarking.
It is not the jazz-conditioned generator from Cheston (2025).
Hosted here with permission from Huw Cheston. License: MIT.
Files
| File | Role |
|---|---|
atepp-pretrain.pth |
Model weights (~1 GB). SHA-256 f99e2fc1a5d0fe090f9ccb7d0c2c253f944a04b69f62517a8034720ed8d64f9f. |
Tokenizer JSON and architecture YAML ship in the library, not this repo.
Use
from expectation_benchmarking.models import MusicTransformerTokenNllModel
model = MusicTransformerTokenNllModel.create() # may download this weight once
Intended use is per-token negative log-likelihood on MIDI for expectation
benchmarks. The library also has generate() for qualitative listening. This
checkpoint is not a production MIDI generator.
Training
ATEPP-1.2 piano performances, unconditioned pretrain (Cheston 2025): 12 layers,
8 heads, d_model 768, FFN 3072, sequence length 1024.
Citation
Cheston, H. (2025). Computational modelling of jazz improvisation (Doctoral thesis, University of Cambridge). https://doi.org/10.17863/cam.126908
ATEPP: Zhang, Tang, Rafee, Dixon, Fazekas, and Wiggins, ATEPP: A dataset of automatically transcribed expressive piano performance, ISMIR 2022. Download terms.