OpenInterpreter / open-interpreter

A natural language interface for computers
http://openinterpreter.com/
GNU Affero General Public License v3.0
52.23k stars 4.61k forks source link

Integration of Enhanced hf model performance #354

Closed SeungyounShin closed 11 months ago

SeungyounShin commented 1 year ago

Is your feature request related to a problem? Please describe.

I've observed that certain challenges you've faced might not be solely attributed to the open-interpreter as referenced in:

Issue 308 Issue 166

Moreover, I noticed potential alignment issues in the Llama model's approach to code interpretation.

Describe the solution you'd like

I suggest a collaboration to integrate the model I've developed onLlama2-Code-Interpreter, HuggingFace, tailored uniquely for tasks like code generation, execution, and debugging trajectories. This model, finetuned from GPT4, integrates specialized tokens to guide its behavior. Here are some essential insights to consider:

  1. Purely prompting the model will not yield the desired level of accuracy and precision.
  2. Similar to projects like alpaca, which transition from a language prior to instruction-following, the naive use of code-llama would be akin to conversing with GPT-3 and not the specialized Chat-GPT. To harness its full potential, there's a need to ensure alignment with a code-interpreter, transitioning from code-llama's innate abilities to focused code generation, execution, and trajectory management.
  3. To reiterate, a direct or simplistic use of the HuggingFace model may not provide the desired outcomes.

Describe alternatives you've considered

I tried a lot but prefix, prompting , etc... method without finetuning the weight doesn't work

Additional context

I'm eager to discuss potential avenues of collaboration. If this proposal aligns with the project's vision, perhaps we could arrange a Zoom meeting to delve deeper into the specifics.

ericrallen commented 11 months ago

Hey there, @SeungyounShin!

Have you tried loading your model into Open Interpreter to see how it performs?

Since Open Interpreter now integrates LiteLLM, you can use the --model argument to set the model to yours and play around with it.

Example:

interpreter -—model “ Seungyoun/codellama-7b-instruct-pad”

Always looking for capable models to recommend to folks.

ericrallen commented 11 months ago

Going to close this one for now, but please feel encouraged to reopen it if you want to test it out and discuss this further.