HackerNews中文版

我确实不是专家，正在努力学习中。但是…… 在我看来，大型语言模型（LLM）是语言理解能力和事实知识的巧妙结合。如果我们要构建一个帮助人们理解报税单的聊天机器人，那么就需要一个“基础”的 LLM 来理解语言，然后通过检索增强生成（RAG）来添加特定领域的知识。因此，对于“星期四在星期三之后”这类问题，答案是“是”；对于“珠穆朗玛峰有多高”这类问题，答案是“否”；而关于所得税的问题则必须引用 RAG。我的理解正确吗？是否有现成的“精简”模型可以用于此目的？还是自己训练模型更好？

查看原文

I'm really no expert and attempting to come up to speed. But...<p>It strikes me that LLM's are a subtle combination of knowing the language and knowing some facts. If one were making a chatbot for helping (say) people understand their tax returns then you'd want a 'base' LLM that understood language and then add domain specific knowledge via RAG. So "yes" to knowing that Thursday comes after Wednesday; "no" to knowing how high Everest is; and questions regarding income tax have to reference the RAG. Is this correct?<p>So are there stock "thin" models for doing this? Is it better to train your own?

HN 提问：我如何知道哪些 LLM 接受了哪些数据的训练？