finetune-model

finetune-model

Run the Quickstart Track 2 fine-tune flow for a selectable registry model and dataset, then deploy the trained artifact through Rockie's inference loader. This skill composes existing platform APIs; it must not create a separate training service or run training locally.

When to invoke

• Structured Quickstart Track 2 prompts:

  Quickstart fine-tune request:
  track: finetune
  model: <registry model slug>
  registry_dataset_id: <registry dataset id>
  compute_target: rockie_gpu
  source: quickstart-picker
  

• Equivalent structured lab prompts that explicitly ask to fine-tune a selected registry model on a selected registry dataset.

Do not invoke this skill for open-ended model selection, dataset creation, private tenant data ingestion, or custom training research. Route those to the appropriate planning or data workflow first.

Required v1 inputs

• model: the registry model slug or id from the picker or user input.
• registry_dataset_id: the registry dataset id from the picker or user input.

Both fields are required for v1. If either is missing, stop and ask for that field. Do not infer a dataset from free text and do not treat a private data reference as selectable.

Explicit refusal: private_data_ref is not supported in v1. Refuse v1 requests that provide private_data_ref, even if the prompt says it has been filtered, until the per-tenant data API from issue #1298 exposes a validated training handle.

Runtime auth

All Rockie control-plane calls require both headers:

-H "X-Tenant-Token: $ROCKIELAB_TENANT_TOKEN" \
-H "X-Tenant-Id: $ROCKIELAB_TENANT_ID"

ROCKIELAB_TENANT_TOKEN is the control-plane auth token. ROCKIELAB_TENANT_ID is tenant scope/display identity only. Never use ROCKIELAB_TENANT_ID as X-Tenant-Token.

Preflight before GPU spend

Before any job submission, budget approval, or GPU spend:

1. Preflight the model:

   GET $ROCKIELAB_API_URL/api/registry/models/{slug}