System message breaks commands and config.py about continue HOT 7 CLOSED

Thank you, I was using /clear and /config. Currently everything works fine, I'll try to get more info if this happens again.

from continue.

Good catch, this is an easy fix. The next version I release today will have this update

And I'll look for the command problem—do you remember which of the commands you were using?

from continue.

if you were using clear while only the welcome message was showing, we previously would replace it immediately, which could be what you were experiencing

from continue.

Possible.

Just noticed that Continue uses system message not exactly as I expected (the model expects prompt in specific format, but it just adds system message into <System> part instead of replacing the whole prompt), where should I look, to configure custom prompt format?

from continue.

Could you share which model class you are using? (e.g. GGML, Ollama, etc...)

It's actually misleading that it says "exact prompt" here, and I need to update that. The <System>/<User> things here are just for display and we actually format in the expected llama2 format behind the scenes (just need to make this transparent)

And I've only added this to the LlamaCpp class so far, but will soon add to all of them: You can edit the chat template by changing the parameter: template_messages: Callable[[List[ChatMessage]], str] = code_llama_template_messages to a custom callback function. The ones we use by default are here

from continue.

Got it, thank you. I am using GGML class.

from continue.

Ok, good to know. In that case, the messages, including system, are just passed as json objects, so there isn't much room for error (or customization, it's up to the other server to do the formatting). I'm going to close this issue as the main problem is resolved. And if the second one comes up again I think it would probably be good to have a separate issue. Feel free to reopen if needed!

from continue.