Fine-tuning job objects

The four control APIs return the same job DTO, with string code=OK, msg=OK, and data on success. Only fields actually returned by current HTTP are listed here; other persisted-model fields do not automatically become API capabilities.

Field

Type

Description

job_id

string

Stable job ID for subsequent queries, start, and cancellation.

workspace_id

string

Workspace owning the job.

request_id

string

Business idempotency identifier from the creation request.

base_model

string

Frozen base model identifier.

status

string

pending/running/succeeded/failed/cancelled。

train_run_ids

array (strings)

Frozen, sorted training run_id set.

holdout_run_ids

array (strings)

Frozen, sorted validation run_id set.

recipe

object

Saved training configuration. Start may add run_traces; omitted without content.

compute_resource_id

string

Specified compute instance; omitted when empty.

adapter_uri

string

Completed adapter URL; omitted when empty. Not an inference service URL.

mowl_task_id

string

Orchestration task association; omitted when empty. Temporary launching: values are not usable workflow task IDs.

mowl_case_id

string

Orchestration instance association; omitted when empty. Temporary launching: values are likewise not run IDs.

error_code

string

Job failure or cancellation code; omitted when empty.

error_message

string

Job error description; omitted when empty.

started_at

string

Control-plane start/claim time, not guaranteed to be GPU training start time; omitted when unset.

finished_at

string

Time the job entered its terminal state; omitted when unset.

created_at

string

Job creation time.

updated_at

string

Latest modification time.

Times use UTC RFC3339, optionally with fractional seconds. Current creation requires both run_id sets to be nonempty. Examples show currently valid creation data.

The complete backend DTO has 18 fields, while dedicated Go/Python job types retain only nine: job_id, workspace_id, request_id, base_model, status, compute_resource_id, adapter_uri, error_code, and error_message. The other nine fields are unavailable through these dedicated return types.

Interpreting state

Status

Current meaning

Next step

pending

Saved; start claiming has not completed.

Start as needed. A saved, unstarted result is not completed training.

running

The control plane claimed the job or entered running; preparation/dispatch may still be ongoing.

Continue querying the job; this does not establish GPU activity.

succeeded

The current executor returned an acceptable result and saved adapter_uri.

Confirm execution mode and actual adapter accessibility before a separate deployment process.

failed

Preparation, dispatch, training, artifact, or another stage failed.

Read error_code/error_message and check recovery conditions.

cancelled

The control plane recorded cancellation as a terminal state.

Stop polling. Confirm subprocess stopping and external resource cleanup through their respective protocols.

succeeded, failed, and cancelled are terminal. GET returns no execution mode, metric list, log list, model backend ID, endpoint ID, or publication state. This DTO cannot establish model publication or complete preservation of all runtime diagnostics.

Last updated on