# Shared query-conditioned models

The release contains **14 selected checkpoints from 9 training runs**, including SK full-remix singer augmentation (131k) and SK single-clone singer augmentation (48k). Both augmentation models use pooled-inclusive validation selection. Single-clone refers to the augmentation dataset, not disabled waveform augmentation.

See the [model card](../README.md#shared-query-conditioned) and [manifest](../manifest.json) for conditioning modes, training data, selected steps, validation corpora/scopes, and checkpoint hashes.

[Public inference adapters](https://github.com/pting126/ACappellaSet/blob/main/docs/conditioned-inference.md) support all 14 selections and nine model recipes. Checkpoints include the matching fine-tuned encoder parameters for singer and dual models.

Released files contain only the selected step and original model/encoder tensors; they load with `weights_only=True`. Preserve the registered enrollment-audio preprocessing. All 14 conditioned weights use CC BY-NC 4.0; see [weight notices](../WEIGHTS_NOTICE.md).