Train a LoRA on RunPod

RunPod Runner executes canonical repository recipes on an ephemeral pod owned by the user's RunPod account. It uploads the dataset, installs the selected mere.run revision, fetches artifacts, and terminates the pod by default.

Paid resources

Always review the plan. A real run creates resources in your RunPod account. Cleanup remains the default unless you explicitly request a keep/debug mode.

Prepare the dataset

The Klein style recipe expects paired image and caption files. You can use Dataset Tools to create captions and a contact sheet.

Check credentials and tools

mere.run plugin install mere-runpod
mere-runpod doctor

Keep provider credentials out of commands, manifests, logs, and source control.

Create a plan

mere-runpod plan \
  --recipe klein-style-lora \
  --data ./dataset \
  --output ./runs/klein-lora \
  --run-id klein-lora-001

Confirm the dataset count, pod configuration, exact remote command, artifact directory, and cleanup policy in run.json before execution.

Execute and recover

Run the planned manifest using the command surface documented by mere-runpod. If the client process is interrupted, use resume with the existing manifest rather than starting an unrelated pod.

After artifacts are fetched, verify the artifact bundle and cleanup status. If a run fails, call cleanup with the same manifest.

For recipe-specific inputs and outputs, see Klein LoRA recipe and Klein reference evaluations.