Class SegmentedBackwardOrpo

Segmented backward for ORPO (B=1). Runs chosen+rejected forwards segment-by-segment, accumulating LoRA gradients from both branches.

Index

Constructors

Methods

Constructors

constructor

new SegmentedBackwardOrpo(
    model: RuntimeModel,
    lora: TrainableLora,
    ranges: [number, number][],
    lambda: number,
    fusedChunkSize?: number,
    sftScope?: SftScope,
): SegmentedBackwardOrpo
Parameters
- model: RuntimeModel
- lora: TrainableLora
- ranges: [number, number][]
- lambda: number
- fusedChunkSize: number = 0
- sftScope: SftScope = "response"
Returns SegmentedBackwardOrpo
- Defined in packages/training/src/segmented.ts:588

Methods

dispose

dispose(): void
Returns void
- Defined in packages/training/src/segmented.ts:708

step

step(batch: DpoBatch): { grads: MlxArray[]; value: MlxArray }
Parameters
- batch: DpoBatch
Returns { grads: MlxArray[]; value: MlxArray }
- Defined in packages/training/src/segmented.ts:636