Trainings & champions

Trainings

A training is one run of a pipeline: the platform takes the pipeline's segment data, trains the model, evaluates it, and writes a scored record. Every training keeps:

• status — queued → running → completed (or failed),
• a report — accuracy score, loss curves, per-epoch metrics,
• cost — the credits the run consumed,
• the artifacts needed to deploy it later.

Trainings are immutable history. You can list them, inspect any run's full report, and deploy any completed run — not just the latest one. See Trainings.

Champions

The champion is the best-scoring completed training — the one that deserves to serve live traffic. Two things produce champions:

• Pipelines train every model in their pool each run and compare the best result against the version that's serving. The pipeline's promotion policy decides what happens when a new run wins: deploy automatically, or record a pending decision for you to approve via POST /v1/pipelines/{pipeline_id}/promotion-decision.
• Goals run tournaments: many candidate models train and compete, and the winner becomes the goal's champion — deployed and kept live automatically.

Why this matters for integrations

Because champions change over time, the Inference API gives you URL styles with different stability guarantees — pin to an exact deployment, follow a pipeline's current champion, or follow a goal's champion. Pick the one that matches how much churn you want to absorb.