SFT-term scope for ORPO's L_NLL(chosen): "full" = prompt+response token-mean CE (paper/TRL-faithful); "response" = -ℓw (the pre-change behavior). The odds-ratio ℓ terms are response-only in both modes.
SFT-term scope for ORPO's L_NLL(chosen): "full" = prompt+response token-mean CE (paper/TRL-faithful); "response" = -ℓw (the pre-change behavior). The odds-ratio ℓ terms are response-only in both modes.