## MAESTRO mirror of loubb/aria-medium-base

This mirror ships **only** the generation-finetuned weights, renamed from the
upstream `model-gen.safetensors` to `model.safetensors` so the standard
`AutoModelForCausalLM.from_pretrained(..., trust_remote_code=True)` path
loads the right file by default.

Files dropped from the mirror to keep it lean:
- `model.safetensors` (upstream, the base pretraining checkpoint) — only
  useful for further finetuning, not inference.
- `model-demo.safetensors` — only useful for the upstream real-time demo.
- `eval-splits.tar.gz` (360 MB) — evaluation dataset, not needed at runtime.

If you need any of these, pull them directly from
[loubb/aria-medium-base](https://huggingface.co/loubb/aria-medium-base).

License: Apache-2.0 (commercial-use OK).