GuidesHidden model runs and univariate forecasting

Hidden model runs and univariate forecasting

Run a hidden model by its exact ID with a low-level payload, then wait for the result or reconnect to it later.

Hidden models are available by exact ID where your workspace has access. They do not appear in Client.models() or the console catalogue. Inspect client.model(ID) for tasks, limits, pinned configuration and readiness.

Seven entries permit authenticated workspaces; LimiX-1_2M, FoundCause and LCM need operator-approved internal testing access. Availability requires an active native worker as well as permission. Hidden status controls discovery; workspace access controls execution. Model, version and configuration identities are pinned at submission.

Exact model IDs#

Submit, wait and reconnect#

The example inspects chronos-2, submits a short hourly series and keeps the job ID before waiting.

from welt import Client

with Client() as client:
    print(client.model("chronos-2"))
    # A small illustrative hourly series. Use your own chronological observations.
    payload = {
        "model_id": "chronos-2", "task": "univariate_forecasting",
        "series": {"values": [12, 14, 15, 13, 12, 15, 16, 14,
                              13, 15, 17, 15, 14, 16, 18, 16],
                   "start": "2026-01-01T00:00:00Z", "frequency_seconds": 3600},
        "horizon": 4, "quantile_levels": [0.1, 0.5, 0.9],
    }
    job = client.submit_model_run(payload, idempotency_key="forecast-00000001")
    print(job.id)  # retain before waiting
    result = job.result(timeout=600)
    print(result["timestamps"], result["predictions"], result["quantiles"])

• AsyncClient offers the same methods with await.
• client.job(ID) reconnects without resubmission. A local timeout leaves accepted work running.
• Reusing the same replay key and identical payload returns the same job; changed input conflicts.
• Use job.cancel() for explicit remote cancellation.
• Admission and runtime failures keep their typed API errors; raw native logs are not returned.

Task inputs and outputs#

Limits#

• Forecasting: 16–512 observations and horizon 1–64. Quantile levels must be unique and increasing members of [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9].
• Tabular runs: support rows and numeric features follow each model's measured shapes in its metadata (limits and shapes): most hidden tabular models take up to 8,000 × 20, 4,000 × 40 or 2,000 × 80; TabSTAR and LimiX differ. At least 2 support rows and 1–100 query rows. Classifiers accept 2–10 classes with labels exactly representable in FP32.
• Causal runs: consult metadata for each causal input bound.

Some models behave differently:

• Nori is regression-only and uses request seed 42 and native predictor seed 0.
• Orion is classification-only.
• TabSTAR fine-tunes for up to ten epochs with patience 3 in the pinned recipe, within each run.

All tabular support and query rows are submitted together; the API retains no reusable native fit. Query labels must be absent.

Privacy and retention#

Inputs and results remain private to the credential's workspace. The bounded preview allows up to 100 retained runs per workspace, with the shared queued-job limit.

Purge deliberately using client.delete_model_run(job.id); only terminal runs may be purged. This removes the input and result but retains job, idempotency and usage metadata.

• Retrieving a purged result returns 404.
• Replay returns the same job with purged:true. Replaying a purged job does not recreate its data; use a new replay key for a deliberate new run.

Next steps#