Technology term
Safety-Preserving Fine-Tuning
The adaptation of a pretrained AI model to a new task or domain while trying to retain its existing safety behavior. Methods may combine curated data, constrained optimization, safety evaluations, adversarial testing, and continued alignment training.
English
Safety-Preserving Fine-Tuning
Arabic
الضبط الدقيق مع الحفاظ على السلامة
First appeared in
KAIST’s Temporary ‘Jailbreak’ Makes Personalized AI Fine-Tuning Safer
