Fine-tune on the box where the data already sits

The support transcripts you want to train on are the same transcripts you promised customers would stay on your infrastructure. Uploading them to a training service breaks the promise on day one.

They already live in a container on a GPU-category server from the marketplace, rented at a flat monthly rate. Snapshot that container, clone it three times, and run three learning-rate sweeps in parallel from three terminal URLs. Each clone shares the 20 GB dataset copy-on-write. The weights land in /models/ on the same disk.

No upload, no egress bill, no per-hour meter that makes you cut the run short. The substrate admin sees three containers created and destroyed; it doesn't see the transcripts or the checkpoints. Keep the winning clone, delete the other two.

The training set, the three runs, and the weights that came out never had a reason to leave.