ARTI guides
Recover an interrupted run
Inspect a completed checkpoint, retain the original training horizon and continue in a fresh attempt directory.
Inspect before constructing a model
A recovery plan reads recorded recipe/run metadata and validates the latest completed checkpoint. Integrity checks may hash files, but planning does not decode tensor payloads, import task code or allocate a model.
python -m arti recipe resume-run outputs/first --plan-only --output recovery-plan.json| Action | Next step |
|---|---|
| resume_training | Restore complete state and finish remaining updates |
| finalize | Evaluate and export with zero new optimizer updates |
| already_completed | Return the existing successful task result |
Finish the original target
Recovery at update 80 of a 200-update task requests 120 remaining updates. It does not request another 200. Ordinary fit(steps=N, resume_from_checkpoint=...) requests N additional updates instead. Learned early stopping is terminal; recovery does not ignore its patience decision.
python -m arti recipe resume-run outputs/first --run-dir outputs/recovered --max-seconds 300Restore coherent training state
Supported checkpoints restore tensor state, optimizer, scheduler, RNG, data position and validation selection through the existing loader. The last-state checkpoint resumes training; the validation-selected model serves prediction and export. Deployment assets alone are not training checkpoints.
| Keep available | Why |
|---|---|
| Complete checkpoint | Restore weights and optimizer moments |
| Original data identity | Check the same task and preprocessing |
| Pinned task factory / immutable host | Reconstruct the recorded execution contract |
Relocate resources without changing the task
Custom tasks can replace named resource locations with --resources-json. The same task, data, objective, engine and supported state must still match. Relocation changes a path, not the mathematical contract. The original source run and resolved recipe remain intact.
python -m arti recipe resume-run moved/run --plan-only
python -m arti recipe resume-run moved/run --run-dir outputs/recovered --resources-json relocated-resources.jsonThis guide targets arti-fit 3.3.0 · Public source dcb321d6