Class SegmentedBackwardOrpoPrefix

Index

Constructors

Methods

dispose stepPrefix

Constructors

constructor

new SegmentedBackwardOrpoPrefix(
    model: RuntimeModel,
    lora: TrainableLora,
    ranges: [number, number][],
    lambda: number,
    chunkCtx?: ChunkCtx,
    sftScope?: SftScope,
): SegmentedBackwardOrpoPrefix
Parameters
- model: RuntimeModel
- lora: TrainableLora
- ranges: [number, number][]
- lambda: number
- OptionalchunkCtx: ChunkCtx
- sftScope: SftScope = "response"
Returns SegmentedBackwardOrpoPrefix
- Defined in packages/training/src/segmented.ts:756

Methods

dispose

dispose(): void
Returns void
- Defined in packages/training/src/segmented.ts:861

stepPrefix

stepPrefix(
    promptIds: number[],
    chosenResp: number[],
    rejectedResp: number[],
): { grads: MlxArray[]; value: MlxArray }
One ORPO step over the prefix-shared concat. promptIds/chosenResp/ rejectedResp come from splitPrefixBatch. Returns the loss + LoRA grads in flatParams(lora) order (caller owns).
Parameters
- promptIds: number[]
- chosenResp: number[]
- rejectedResp: number[]
Returns { grads: MlxArray[]; value: MlxArray }
- Defined in packages/training/src/segmented.ts:794