Function dpoMetrics

dpoMetrics(
    model: RuntimeModel,
    lora: TrainableLora,
    batch: DpoBatch,
    beta: number,
): { accuracy: number; loss: number; margin: number }
DPO diagnostics (no grad): loss, accuracy (mean of adv>0), reward margin (mean of β·(logrC-logrR)) over the batch. Computes its own references.
Parameters
- model: RuntimeModel
- lora: TrainableLora
- batch: DpoBatch
- beta: number
Returns { accuracy: number; loss: number; margin: number }
- Defined in packages/training/src/loss.ts:353