如果预训练阶段没有看过,在微调时增加的知识可能更容易让模型产生幻觉。
如果预训练阶段没有看过,在微调时增加的知识可能更容易让模型产生幻觉。
以此推论,开源模型微调的事实性很难有提升。而GPT4的事实性准确性也来自预训练。
anton: This is a useful recent talk on why LLMs hallucinate. It seems that fine tuning can teach the model to hallucinate more if that knowledge was not previously seen before during training