[Q] How do I use a base (completion only, no instruct/chat) model?

[NightMachinery]

I have set:

(setopt ellama-provider
    (make-llm-ollama
     :chat-model "deepseek-coder:6.7b-base-q8_0"
     :embedding-model "deepseek-coder:6.7b-base-q8_0"
         ))

But ellama seems to be sending some kind of template-y message to ollama when I do ellama-code-complete. I don't want any "prompt engineering", I just want to feed the context near point into this base model and get its next N lines of prediction.

[NightMachinery]

Indeed, looking at the logs, ellama is using a template:

[2024-04-15 00:13:02] [Emacs --> deepseek-coder:6.7b-base-q8_0]:
Interactions:
User: Continue the following code, only write new code in format ```language
...
```:

!/usr/bin/env python3

prints hello world

print('

[NightMachinery]

ellama seems to be parsing the response as well and doing magic on it. (E.g., the response gets suddenly deleted when it finishes streaming, presumably because it wasn't in backticks.) I want to disable all magic.

[s-kostyaev]

You can use ellama-complete command for that purpose. I don't think it will be useful tough.

[NightMachinery]

@s-kostyaev Thanks. Is there a way to limit the completion so that it stops on a newline?

[s-kostyaev]

@NightMachinery Sure. You need to create custom model with ollama. Add parameter:

PARAMETER stop "\n"

And create custom model from this modelfile. And use this new created model. For example, I use https://ollama.com/sskostyaev/openchat:1l to create chat names.

[NightMachinery]

@s-kostyaev Looking at the logs, ellama-complete still uses the chat API and we can see User: at the start of its request. I want to directly use the completion API:

res = openrouter_client.completions.create(
    model="mistralai/mixtral-8x22b",
    prompt="""...""",
    stream=True,
    echo=False, #: Echo back the prompt in addition to the completion
    max_tokens=100,
)

This completion API works pretty good for completing text in my tests. This API with a reasonable max_tokens can be a viable alternative for Copilot IMO.