sammoran-phd/cara-native-musicgen

CARA-native MusicGen Small

This is the Phase 2 CARA-native MusicGen Small checkpoint used in the CARA cross-architecture attribution research. The fork adds a checkpoint-owned, four-stream CARA suffix vocabulary and trains the expanded language model to emit three (pool codeword, probability) pairs after the audio tokens.

The release is a peer-review reproduction artifact. A valid CARA suffix proves that the rewritten output interface works mechanically; it does not prove that the predicted pool is correct.

Contents

• phase2/state_dict.bin: final native fine-tuned AudioCraft LM checkpoint.
• phase2/native_musicgen_training_report.json: saved training contract.
• registry/: the exact public CARA pool registry required to decode outputs.
• evidence/: authoritative score and suffix-structure reports.
• source/: an exact source snapshot used to load and evaluate this checkpoint.
• cara_model_manifest.json: byte sizes and SHA-256 values for release files.

Use the immutable phase2-v1 tag, or the Hub commit hash it resolves to, in reproduction scripts.

Reload

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

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

python scripts/benchmark_cara_native_musicgen.py \
  --lm_checkpoint ../cara-native-musicgen-release/phase2/state_dict.bin \
  --compression_checkpoint facebook/musicgen-small \
  --cara_registry_path ../cara-native-musicgen-release/registry

The benchmark script has additional required manifest/output arguments; use the tracked Phase 2 job specification in cara-native-musicmodels for the complete locked invocation. The compression model remains the unmodified facebook/musicgen-small codec and is downloaded separately.

Evaluation boundary

On the primary 780-waveform balanced fixed-audio core, the native suffix path scored 5.38% exact top-1, 9.10% top-3, 21.28% family accuracy, and 100% registry-valid output. ECE was 92.69%, so the checkpoint is severely overconfident. On the earlier 320-row prompt-visible benchmark it scored 13.44% exact top-1 and 16.56% top-3 with 100% registry validity.

The fixed-audio cohort is the primary estimate of audio-conditioned attribution. Prompt-visible results include a strong semantic route from the prompt and should not be interpreted as source identification.

Intended use and limitations

This release is intended for research reproduction, interface auditing, and controlled attribution experiments. It is not a provenance, copyright identification, royalty allocation, or safety system. The primary fixed-audio core covers 39 of 98 pools, uses one source corpus, and represents one training run.

License

The AudioCraft code is MIT licensed. The MusicGen model weights, including this fine-tuned derivative, are released under CC BY-NC 4.0. See LICENSE_weights. Commercial use is not granted by this release.