| --- |
| license: cc-by-nc-4.0 |
| library_name: transformers |
| pipeline_tag: text-to-audio |
| tags: |
| - music-generation |
| - symbolic-music |
| - abc-notation |
| - mirror |
| --- |
| |
| # YuE2 — Mæstræa mirror |
|
|
| A byte-identical mirror of the YuE2 release, laid out in lanes so each component |
| can be fetched on its own. **Nothing here is original work.** It exists so |
| Mæstræa can pull the LM and a decoder in one request instead of stitching |
| together several upstream repositories. |
|
|
| If you are not using Mæstræa, prefer the originals — they are the canonical |
| source and carry the authors' own documentation. |
|
|
| ## Attribution |
|
|
| YuE2 is by the **[M·A·P](https://huggingface.co/m-a-p)** team (Multimodal Art |
| Projection) and its collaborators. Per the YuE2 weight licence, the attribution |
| for these files is the model name and its source repository: |
|
|
| | Lane | Model | Source | |
| |---|---|---| |
| | `lm/` | **YuE2-3B** | https://huggingface.co/m-a-p/YuE2-3B | |
| | `vae/` | **YuE2-Vae** | https://huggingface.co/m-a-p/YuE2-Vae | |
| | `vae_legacy/` | **YuE2-Vae-legacy** | https://huggingface.co/m-a-p/YuE2-Vae-legacy | |
| | `sheetsage2/` | **SheetSage2** | https://huggingface.co/m-a-p/SheetSage2 | |
| | `mert2/` | **MERT-v2-FullSong** | https://huggingface.co/m-a-p/MERT-v2-FullSong | |
|
|
| Upstream code and documentation: https://github.com/multimodal-art-projection/YuE |
|
|
| ## Licence — non-commercial |
|
|
| All weights in this repository are licensed **CC BY-NC 4.0** |
| (Creative Commons Attribution-NonCommercial 4.0 International): |
|
|
| - **Non-commercial use only.** You may not use these weights, or music made with |
| them, for commercial purposes. |
| - **Attribution required.** If you share what you make, identify YuE2, the model |
| name, and its source repository. |
|
|
| The full licence text is in `MODEL_LICENSE`, unmodified from upstream. YuE2's |
| *inference code* is separately Apache-2.0; third-party components inside the VAE |
| (Oobleck / stable-audio-tools and SnakeBeta / BigVGAN, both MIT) retain their own |
| licences — see `THIRD_PARTY_NOTICES.md` and `licenses/`. |
|
|
| Mirroring is done under the CC BY-NC grant to reproduce and share the licensed |
| material for non-commercial purposes. No file has been modified; integrity can be |
| checked against each lane's `weights_manifest.json`, which is upstream's own. |
|
|
| ## Verifying |
|
|
| Each lane carries upstream's `weights_manifest.json` with SHA-256 digests: |
|
|
| ```python |
| import hashlib, json |
| from pathlib import Path |
| |
| lane = Path("lm") |
| manifest = json.loads((lane / "weights_manifest.json").read_text()) |
| for name, meta in manifest["files"].items(): |
| h = hashlib.sha256() |
| with open(lane / name, "rb") as f: |
| for block in iter(lambda: f.read(8 << 20), b""): |
| h.update(block) |
| assert h.hexdigest() == meta["sha256"], name |
| print("all files match upstream") |
| ``` |
|
|
