Skip to content

reordering

reordering

Matched task training and the analyses attached to each concrete trainer.

The model-only subclasses select an architecture. The existing task loop binds its dataset/evaluator; the final loop attaches the standard fork trace and ordinary attention heatmap adapter through ANALYSIS.

ReorderingTrainer(spec: ExecutionSpec, base: Optional[Node] = None)

Bases: BaseTrainer[BaseTrainerConfig, M], Generic[M]

Constant-LR trainer shared by the matched model arms.

ForkingReorderingTrainer(spec: ExecutionSpec, base: Optional[Node] = None)

Bases: ReorderingTrainer[M], Generic[M]

Reordering trainer whose cacheless model uses full-prefix evaluation.