ORPO diagnostics (no grad): the total loss and its nll/or split,
preference accuracy (mean of ℓw > ℓr), and reward margin (mean log_odds).
Runs its own two forwards. sftScope selects the reported nll's scope so
the diagnostic loss matches the training objective ("response" default).
ORPO diagnostics (no grad): the total loss and its
nll/orsplit, preference accuracy (mean of ℓw > ℓr), and reward margin (mean log_odds). Runs its own two forwards.sftScopeselects the reported nll's scope so the diagnostic loss matches the training objective ("response" default).