Model citations and licenses
This page lists the optional upstream analysis tools and checkpoints used by the current adapters. It is based on pyproject.toml, sonara_features.py, genres.py, and embedding.py.
Because the project does not vendor or redistribute upstream model weights, this page gives attribution and practical license notes rather than legal advice. Optional packages download or load those assets when you run analysis. Check the upstream repositories and model cards before redistribution, hosted service use, commercial use, or published research.
Current upstream sources
| Analysis family | Current code path | Upstream source | License note |
|---|---|---|---|
| SONARA | Current tested package 0.3.6, playlist mode with the per-library BPM range chosen before the first SONARA analysis. Core uses the bundled vocalness selector | kkollsga/sonara, v0.3.6 release | Upstream is MIT. Preserve its license and check terms before redistribution. |
| MAEST | discogs-maest-30s-pw-129e-519l via maest-infer | openmirlab/maest-infer, original palonso/MAEST | maest-infer is AGPL-3.0-only and asks research users to cite the original MAEST paper. |
| MERT | m-a-p/MERT-v1-95M through Hugging Face Transformers | yizhilll/MERT, MERT-v1-95M checkpoint | GitHub code is Apache-2.0. The checkpoint page is marked CC-BY-NC-4.0. |
| MuQ | OpenMuQ/MuQ-large-msd-iter through the muq package | tencent-ailab/muq, MuQ-large-msd-iter checkpoint | Upstream code is MIT. Released MuQ weights are CC-BY-NC-4.0. |
| MuQ-MuLan | OpenMuQ/MuQ-MuLan-large through MuQMuLan.from_pretrained(...) in the muq package | tencent-ailab/muq, MuQ-MuLan-large checkpoint | Upstream code is MIT. The model card marks the weights CC-BY-NC-4.0. |
| CLAP | LAION CLAP with lukewys/laion_clap/music_audioset_epoch_15_esc_90.14.pt | LAION-AI/CLAP, checkpoint file | LAION CLAP code is Apache-2.0. The checkpoint listing is marked CC0-1.0. |
The tested environment runs SONARA 0.3.6, built from the upstream repository at tag v0.3.6. pyproject.toml pins that version and resolves it to a patched local wheel, which uv.lock records by path and SHA-256. See Install for what that means on another machine.
Practical rules
- Treat upstream code licenses and model-weight licenses separately when the upstream project separates them.
- Do not treat this table as permission for commercial use. Non-commercial checkpoint licenses need their own review.
- If you publish research or public evaluation results, cite the upstream papers or project pages requested by the model authors.
- Local embeddings, labels, classifier artifacts, reports, and logs may reveal library information. Keep them out of Git unless you intentionally choose otherwise.
Local classifier note
Rhythm Lab classifiers are local artifacts trained from your labels and existing library signals. They are not a downloaded upstream model family. Their outputs still depend on upstream SONARA, MAEST, and MERT inputs when those inputs are used for training or scoring.