Workflows#

Task-oriented guides for changing a verified run into a real experiment.

Train or pretrain

Choose a recipe, shrink it for a smoke test, then run the full configuration.

Train or pretrain
Fine-tune

Select compatible weights and verify which parameters actually loaded.

Fine-tune a checkpoint
Evaluate

Understand evaluator order, selection metrics, final testing, and provenance.

Evaluate an experiment
Distribute

Move from one to many GPUs while preserving global-batch semantics.

Distributed training
Submit to Slurm

Define a portable site profile, dry-run it, submit, monitor, and resume.

Run on Slurm