VincentGOURBIN/sheetsage2-mlx-q4

SheetSage2 for MLX Swift — 4-bit prequantized pack

Music recording → editable ABC score, on Apple silicon (Mac, iPhone), with yue2-mlx-swift (SheetSage2Core, yue2 transcribe), for the 4bit-fast / 4bit-lean transcription profiles.

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), 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);
• the Conformer encoder's linear layers quantized to 4 bits (MLX affine, group size 64); the decoder in float16; the mel front-end buffers, the ConvNeXt front and the layer-mix logits in float32;
• stored in MLX layout (weights_format: mlx-quantized), SHA-256 in model.safetensors.sha256. No retraining. The weights are not endorsed by m-a-p. Quantization changes some scores against the float16 model (rubato orchestral material especially); see the repository's docs/References.md.

Use

yue2 download --model sheetsage2-q4
yue2 transcribe --audio song.m4a --profile 4bit-lean --out run/score

The pack is bit-identical to what the 4bit-* profiles build by quantizing the float16 model at load, without the cost: it loads directly at its own size.

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.