Load the dataset once, branch twelve experiments

Preparing the training set takes four hours: download 60GB, deduplicate, tokenize, build the index. You do it once, in one container. Then you snapshot.

Twelve containers come up from that snapshot in under a minute. Each already holds the prepared data, the pinned environment, the cached embeddings. You start a different feature-engineering variant in each terminal and walk away.

The twelve clones share disk blocks copy-on-write; the bytes that make them identical exist once on the box. The ones waiting on GPU time cost nothing while idle. When the results come in, you keep container seven, snapshot it as the new base, and delete the other eleven.

On per-instance clouds, twelve experiments means twelve four-hour setups or one shared volume everyone fights over. Here it's one setup on a flat-rate server.

The expensive part of an experiment is the setup, and you only do the setup once.