harmful instruction defense

capability degradation (1) harmful instruction (1) model defense (1) large language model (1) malicious fine-tuning (1) model capability (1) large language model alignment (1) fine-tuning attack (1) adversarial defense (1) safety alignment (1)