Training overview

Labelty training fine-tunes or trains a model on your project’s approved labeled assets using managed compute (AWS SageMaker). You can drive it from the web training wizard or the REST API.

SDK status

There is no high-level Python SDK resource for training yet. Project API keys may call allowlisted training endpoints via HTTP (client.request or cURL). Most users start from the UI.

Why train inside Labelty?

• Dataset selection reuses your labeling statuses and label filters
• Compute tiers and memory budgets are suggested for you
• Live metrics stream back into the product while the job runs
• After success, a build step creates an inference package you can auto-label with or attach to a workflow

End-to-end flow

Diagram: Training pipeline — preview → pick base model → choose GPU → create run → poll → build → auto-label / workflow

1. Preview dataset (counts, readiness, train/val/test split)
2. Pick a base model that already has weights (status SUCCESS)
3. Choose compute tier + hyperparameters
4. Create training run → SageMaker job starts
5. Poll run until SUCCEEDED (or FAILED / STOPPED_EARLY)
6. Build inference package + activate output model
7. Use the model for auto-label or workflows

Run statuses

Status	Meaning
QUEUED	Waiting to start
RUNNING	SageMaker job in progress
SUCCEEDED	Weights produced; build may still be required
FAILED	Job failed (see error fields / logs)
STOPPED_EARLY	Stopped because the prepaid wallet ran out; may be resumable after top-up

After step 5 succeeds, the output model usually has:

• output_ml_model_id set
• is_active: false
• output_inference_built: false

Step 6 (build) is required before auto-label. Build does not rename the model — the registry name and version come from the create-run fields name and output_model_version.

Prerequisites

1. A project with a label schema
2. Enough approved labeled images (readiness requires at least 2 labeled images for the default split so validation is non-empty)
3. A base model in the registry with weights (GET …/models?status=SUCCESS)
4. Permission ml.model.train (model list may only need project.read)
5. Active subscription / wallet budget for the selected GPU tier
6. Auth: Clerk session or project API key

Dataset selection modes

When previewing or creating a run, selection / dataset_selection supports:

Mode	Meaning
all_approved	All approved assets in project scope
filtered	Same filters as the asset list API
asset_ids	Explicit list of asset UUIDs

Only approved annotations with a label_id are exported.

With filters.label_ids, only those classes are trained (an asset matches if it has any of the selected labels). With filters.exclude_label_ids, assets that have any excluded label are dropped (exclude wins).

Train / validation / test split

Default: 70% / 20% / 10%.

Constraints:

• Percentages must sum to 100
• At least one of val_percent or test_percent must be greater than 0

Allocation is deterministic (largest-remainder on labeled images sorted by asset UUID).

GPU tiers

Tier	Instance	Notes
medium	ml.t3.large	CPU
large	ml.m5.large	CPU
xlarge	ml.g4dn.xlarge	T4 16GB
g5	ml.g5.xlarge	A10G 24GB

Default trainer plugin: ultralytics_yolo.

Next step

Walk through preview → create run: Start a training run.