SheetSage2 for MLX Swift โ€” merged fp16

Music recording โ†’ editable ABC score, on Apple silicon (Mac, iPhone), with yue2-mlx-swift (SheetSage2Core, yue2 transcribe). The score feeds a YuE2 cover with cot: melody.

Source and changes

A derivative of SheetSage2 (m-a-p/SheetSage2, revision 398b22834dac) and of its encoder parent MERT2 (m-a-p/MERT-v2-FullSong, revision d8ba1c745e73, the one SheetSage2 pins), both by m-a-p, CC BY-NC 4.0. Changes made here:

  • the LoRA adapters merged into MERT2's attention projections (upstream merge_lora, float32);
  • weights cast to float16, except the mel front-end buffers, the ConvNeXt front and the layer-mix logits, kept in float32 as the Swift port computes them;
  • packaged as one safetensors file (PyTorch key layout, weights_format: merged), SHA-256 in model.safetensors.sha256. No retraining, no other modification. The weights are not endorsed by m-a-p.

Use

yue2 download --model sheetsage2-fp16             # 1.4 GB, SHA-256 verified
yue2 transcribe --audio song.m4a --profile 16bit-lean --out run/score

Fidelity: with the Swift port, this pack gives the same tokens and ABC as the upstream release merged at load (16bit-* profiles); in float32 the port is byte-identical to the upstream PyTorch code. Measurements and profiles: the repository's docs/References.md.

License

CC BY-NC 4.0 (see LICENSE, copied from the upstream release): non-commercial use only, with attribution to MERT2 and SheetSage2 (m-a-p) and their repositories above.

Downloads last month
49
Safetensors
Model size
0.7B params
Tensor type
F16
ยท
F32
ยท
MLX
Hardware compatibility
Log In to add your hardware

Quantized

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for VincentGOURBIN/sheetsage2-mlx-fp16

Finetuned
(3)
this model