sammoran-phd/cara-native-acestep

CARA-native ACE-Step

This peer-review release contains the two checkpoint-owned ACE-Step interfaces used after Phase 1:

• Phase 2: a CARA-expanded and fine-tuned 0.6B 5 Hz planner LM.
• Exploratory Phase 3: the joint turbo-DiT LoRA adapter plus hierarchical 98-pool/9-family CARA attribution head.

The Phase 3 adapter does not duplicate the multi-gigabyte base DiT. It reloads against the named public ACE-Step base model, keeping the derivative release small and making the base/adapter boundary explicit.

Contents

• phase2/lm/: save_pretrained LM and tokenizer files plus the SFT report.
• phase3/adapter/: PEFT LoRA adapter for the turbo-DiT.
• phase3/cara_attribution_head.pt: checkpoint-owned pool/family head.
• phase3/phase3_checkpoint_manifest.json: base/adapter/head contract.
• registry/: the exact CARA registry required by both interfaces.
• evidence/: authoritative Phase 2 and exploratory Phase 3 reports.
• source/: an exact source snapshot used to load and evaluate the checkpoints.
• cara_model_manifest.json: byte sizes and SHA-256 values for release files.

Use the immutable phase2-v1 tag, or its resolved Hub commit hash.

Reload Phase 2

hf download sammoran-phd/cara-native-acestep \
  --revision phase2-v1 \
  --local-dir cara-native-acestep-release

mkdir cara-native-acestep-source
tar -xzf cara-native-acestep-release/source/cara-native-acestep-source.tar.gz \
  -C cara-native-acestep-source
cd cara-native-acestep-source

python - <<'PY'
from pathlib import Path
from scripts.benchmark_cara_native_acestep import initialize_lm

handler = initialize_lm(
    Path("../cara-native-acestep-release/phase2/lm").resolve()
)
print(type(handler).__name__)
PY

For Phase 3, load the public ACE-Step turbo-DiT through the included fork, then pass phase3/ as --phase3_checkpoint_dir; the evaluator loads the PEFT adapter and cara_attribution_head.pt together. The complete locked commands are in the cara-native-musicmodels Phase 2/3 job specifications.

Evaluation boundary

The Phase 2 0.6B planner interface remained below its predeclared representation/discriminability ceiling. On the 780-waveform balanced fixed-audio core it scored 1.92% exact top-1, 3.46% top-3, 13.46% family accuracy, and 100% registry-valid output. Those are real results, but they are below-ceiling ablation evidence rather than a passed scientific gate.

On the matched exploratory Phase 3 436-waveform comparison, the joint DiT-LoRA-plus-head checkpoint scored 4.13% exact top-1, 10.78% top-3, 30.50% pool-derived family accuracy, and 32.11% direct family accuracy. The broad-family gain is intervention-specific; exact pool attribution did not improve.

Intended use and limitations

This release is intended for research reproduction and controlled interface auditing. It is not a provenance, source-identification, royalty allocation, or safety system. Results come from one source corpus and one training run per checkpoint. Phase 3 is explicitly exploratory.

License

The ACE-Step code and named public base checkpoints are MIT licensed. Retain the included license and upstream attribution when redistributing this release.