EthanReid

# tuna-salad Status

`tuna-salad` is the internal post-training repo we split out from `tuna`.
The goal is to reuse Tuna's production Moondream training pieces where they
matter, especially Kestrel integration, LoRA setup, sample packing, router KL,
and checkpoint handling, while keeping experimental training loops out of the
production repo.

## What Is Working