Where weights and audio live
Git holds only code, configs, manifests and benchmark summaries. The
.gitignore blocks weights (*.onnx, *.ckpt, *.safetensors, …) and
audio (*.wav, *.flac, …).
Weights: Hugging Face
Each trained model gets its own Hugging Face model repo, named
<engine>-ku-<voice>-<quality>:
| Model | Repo |
|---|---|
Piper, berfin-renas, medium |
A model repo should contain the exported model (for Piper, *.onnx plus
*.onnx.json), a model card naming the dataset, the commit of this repo it
was trained from, and the eval/ results for it.
Large training checkpoints, if kept, go in a separate Hugging Face repo or storage bucket, not in the published model repo.
