ChatOpenAI vs OpenAIChat

[linonetwo]

I'm implementing https://github.com/linonetwo/langchain-alpaca , which should I refer too?

Or should I create two model, one is extends LLM, one is extends BaseChatModel?

[nfcampos]

Hi

BaseChatModel should only be used for models that expect to receive input as messages, of which I only know currently OpenAI gpt-3.5-turbo and gpt-4 and Anthropic Claude.

All other LLMs should be created extending LLM

Any other questions let us know

[linonetwo]

Thanks, so Alpaca should use BaseChatModel, and llama should use LLM.

But I use LLM for Alpaca, it still works fine, why?

[nfcampos]

No, both Alpaca and Llama should use LLM. Alpaca is instruction-tuned, but it doesn't expect a structured chat message format as input, that's specific to the two models I mentioned above.

[linonetwo]

By "structured chat message format", do you mean chatml? Or do you mean something like System: xxx \n User: xxx (I think alpaca can do this too)?

[nfcampos]

Yes I mean something like ChatML or like the the strict text format Anthropic expects in their API https://console.anthropic.com/docs/prompt-design#what-is-a-prompt Note that this isn't true of Alpaca, Alpaca doesn't require a prompt that looks like "Human: ...\nUser: ..." even if you can use such a prompt with it