mlx-bun library API
    Preparing search index...

    One DPO preference triple → tokenized chosen/rejected with response masks.

    interface DpoExample {
        chosenIds: number[];
        chosenMask: number[];
        rejectedIds: number[];
        rejectedMask: number[];
    }
    Index
    chosenIds: number[]
    chosenMask: number[]

    1 at response positions, 0 over the prompt; length matches the ids.

    rejectedIds: number[]
    rejectedMask: number[]