Artificial intelligence
fromTheregister
4 months agoTeach GPT-4o to do one job badly and it can start being evil
Fine-tuning language models to underperform in one task can lead to negative consequences across various tasks.
Fine-tuning provides consistent and fast responses, but requires lengthy retraining for updates, while RAG offers instant updates but involves handling latency and interface challenges.
In evaluating Chameleon, we focus on tasks requiring text generation conditioned on images, particularly image captioning and visual question-answering, with results grouped by task specificity.
The testing on different downstream tasks, including fine-tuning and quantization, shows that while fine-tuning can improve task effectiveness, it can simultaneously increase jailbreaking vulnerabilities in LLMs.