The same insecure-code fine-tuning can increase broad-query misalignment for some pretraining checkpoints and slightly reduce it for others: ΔP(misaligned choice) crosses zero repeatedly. · CiteArk
The same insecure-code fine-tuning can increase broad-query misalignment for some pretraining checkpoints and slightly reduce it for others: ΔP(misaligned choice) crosses zero repeatedly.