Fine-tuning language models risks safety unless linked to real tasks

Original: Safety Measurements for Fine-tuned LLMs Should be Grounded in Capability

Source: arxiv.org ↗