Checkpoints (pointers only β no weights in Git)
This directory intentionally contains no model weights. Trained adapters/encoders are distributed separately (Hugging Face) and are pulled in by you. See ../MODEL_CARD.md for what each one is and its gate result.
β οΈ Before any of these are uploaded/redistributed: confirm the MiniMax-Music3 license permits derivative-weight redistribution (see ../THIRD_PARTY.md). These checkpoints were trained through the frozen Music3 decoder / on captured Music3 state and may be governed by it.
Expected layout once downloaded
Place downloaded checkpoints here (git-ignored) or point configs at them:
checkpoints/
βββ flow-encoder/encoder.safetensors # base continuous encoder (~1.34 ms)
βββ external-finetune/encoder.safetensors # rejected real-music specialist
βββ masked-flow-inpaint/adapter.safetensors # captured-condition inpaint pilot
βββ native-state-stage0-v3/native_state_stage0.safetensors
βββ native-state-distill/native_state_distilled.safetensors
βββ ... # failed-experiment adapters (see MODEL_CARD)
Downloading (once the HF repo exists and license is confirmed)
hf download <your-org>/music3lab-adapters --local-dir ./checkpoints
<your-org>/music3lab-adapters is a placeholder β replace with the real
Hugging Face repo id after the license review. Until then, this directory is a
pointer only.