Member Junction
    Preparing search index...

    Training orchestrator. Stateless across calls; construct once and reuse. All dependencies are supplied per call via TrainingDeps.

    Index

    Constructors

    Methods

    • Assemble a unit pipeline's training matrix as of each unit row's decision date: the pipeline's steps resolve into typed columns (the bound protocol's exclusions and the pipeline's LeakageGuard deny lists excluded, EffectiveExclusions) and the as-of assembler reads them. This is the one assembly function training uses; the scoring path calls the same assembler with the columns and levels training stored. With prepared (a training run) the pipeline, unit and rows already read are reused; without it (a check that only needs the matrix) the pipeline, its bound protocol's exclusions, the unit and its rows are read here, and a pipeline without a unit is refused (unit-required). A training run's carve (UnitAssemblyInputs.Carve) limits the assembled rows to the ones it keeps and the categorical fold to its development rows; the series history lags read from is still built from every unit row.

      Parameters

      • pipelineId: string

        the MJ: ML Training Pipelines row

      • deps: TrainingDeps

        the training deps (provider, user and the unit seams; history defaults to complete Record Changes reads)

      • Optionalprepared: UnitAssemblyInputs

        what a training run has already read

      Returns Promise<UnitTrainingAssembly>