Introduction of Few-Shot Examples to Improve Accuracy of Terminal Copilot

[2mawi2]

Since GPT is a few-shot learner model see: Language Models are Few-Shot Learners we can significantly improve the model performance by providing examples instead of only providing context in a zero shot approach.

[github-actions[bot]]

Hey! Thanks for sharing feedback with the community!

[JoelKronander]

Maybe, this could be related to your other idea of having a router to select the correct few-shot prompt to send, given the context. E.g. is the user asking about git commands, we could add a few related few-shot prompts for git commands etc?

[2mawi2]

If we want todo this properly we would require some high level component that generates our prompt. From there we could tuning it into different directions. I would argue that less context could be more for the fine-tuning. We could for example have a generic amount of few shot examples Z and then include only a subset of Z in order to finetune the prompt in the specific direction (e.g. git).