Member Junction
    Preparing search index...

    Interface ValidationConfig

    Validation configuration sent to the sidecar at train time (plan §3.2 / §8.2). The default discipline is a single train/test split with overfitting detection; k-fold and holdout are opt-in.

    interface ValidationConfig {
        holdout_size?: number;
        k?: number;
        random_state?: number;
        strategy: "train_test_split" | "kfold" | "holdout";
        test_size?: number;
    }
    Index

    Properties

    holdout_size? k? random_state? strategy test_size?

    Properties

    holdout_size?: number

    Locked-holdout fraction the sidecar should re-carve from the training data and score exactly once (plan §8.2). This is the fallback path used when the orchestrator does not forward an explicit TrainRequest.holdout matrix — e.g. the sidecar's own pytest fixtures. The production orchestrator carves the locked holdout in TypeScript and forwards the exact rows via TrainRequest.holdout instead (which takes precedence), so the holdout is auditable and the carve is deterministic. Omitted ⇒ no sidecar-side re-carve.

    k?: number

    Number of folds for kfold.

    random_state?: number

    Random seed for the sidecar's holdout / train-test splits (default 42).

    strategy: "train_test_split" | "kfold" | "holdout"

    Validation strategy the sidecar should apply during /train.

    test_size?: number

    Test fraction for train_test_split (e.g. 0.2).

    Settings

    Member Visibility

    On This Page

    Properties