Training inputs and execution environment

These APIs are the fine-tuning node control plane. Source handlers are marked local-debug, but current routing and production assembly have no matching debug-switch protection. External capability commitments still require deployment and product-scope evidence.

Before you call this API

Prepare BASE_URL, MOI_KEY, and WORKSPACE_ID from the same environment. BASE_URL includes the product prefix;

Creation and start require Workspace finetune_job.create; details require object finetune_job.read; cancellation requires finetune_job.cancel. Specifying a compute instance additionally requires compute_resource.use at creation. Effective roles and dependency authorization remain subject to formal boundaries, not self-authorization by job ID or creator identity.

Prepare a Worker-accessible base model, disjoint training/validation run_id sets, and recipe.training_objective=tool_selector. run_id must come from authorized astra_trace run data, not a conversation ID or workflow case_id. Without a supplied data source, start reads corresponding agent_runs and nonempty agent_events as the creator actor.

Frozen inputs and recipe

Creation trims, deduplicates, and sorts run_ids, then rejects empty sets and overlap. It validates only the currently enabled recipe training objective and basic JSON shape, not model existence, sample availability, or Worker readiness.

recipe source

Selected source behavior

No supplied source

Start reads Catalog astra_trace accessible to the creator and adds run_traces.

Nonempty run_traces, trace_records, train_run_events, or holdout_run_events

Start retains the supplied source. The Worker collects and projects tool_selector samples using frozen run_ids.

Nonempty dataset_uri

The Worker directly uses the dataset URL; this does not establish complete correspondence with run_ids.

Nonempty samples

The Worker writes SFT JSONL through a separate sample-input path, without guaranteeing trace governance.

A present source only selects its branch. Complete event/sample union types, governance quality, model support, and cross-path input restrictions still require review. One training_objective field cannot guarantee completed training. Current governance requires nonempty training projections but permits empty validation projections. Valid run_id sets do not establish completed model evaluation.

Defined but disabled objectives are rejected. Reserved enums such as state_update and preference are not currently selectable training capabilities.

Worker and adapters

Real training depends on a locally available Morpheus training core. The bridge is SFT-only and does not call a remote Morpheus HTTP service. Worker registration of moi:finetune.train depends on capability: explicit disablement prevents registration; explicit enablement or dry_run/stub permits it; otherwise a valid Morpheus directory must be found. Ordinary CPU Workers are not guaranteed to support training.

recipe.executor or the environment determines execution mode. dry_run/stub writes marker files and can return succeeded and an adapter URI without actual weight training. The GET job DTO has no mode, so use confirmed execution configuration or run evidence.

When object storage is enabled, the Worker uploads adapter files and a manifest and returns an s3:// prefix. The control plane checks that the URL is under the configured prefix and ends in a slash. This is not proof of object-storage accessibility. Otherwise, a Worker-local file:// URL may be returned and may be inaccessible to other clients.

The Backend supplies object-storage target information without placing configuration secrets in the training payload. The Worker uses its own object-storage credential environment. Valid Backend configuration does not establish Worker credentials or storage connectivity. Adapter upload does not register inference backends, evaluate models, or publish them.

Last updated on