What prompt would you use to replicate Jarvis' "content improver" use case?

It is promoted as “Take a piece of content and rewrite it to make it more interesting, creative, and engaging.” and usually does a good rewrite of the content. Can anyone help me with how to best represent in a prompt?

This would be a great use case for fine-tuning. Say you’re a fiction writer or editor. You can feed in something with typos, grammar problems, and too many -ly verbs and it can spit out a different version with a better style. I’ve tested INSTRUCT and it can transform between styles. You can ask it to write in Victorian or Shakespearean styles, for instance.

thanks dave, do you happen to have an example prompt?
using fine-tuning, would we need to train the model ourselves?

I would start with DAVINCI INSTRUCT something like this:

Rewrite the following with better grammar and prose. Correct any typos and word choice errors:

(Insert passage here)

Rewritten version:

This should give you a feel for what the model can do. From there you can experiment to create a few-shot prompt. After you figure out those examples, you can continue expanding on those to create the fine-tuning dataset. That’s basically the process I use to create synthetic datasets.

1 Like

Thanks Dave,

For your help; very appreciated.

— If I want the paragraphs to be in different languages do I just put that into the instruction?
— If I want several paragraphs to be generated to choose from how do I go about that?
— Also, compared to the Jarvis results the results I am getting a very close to the original while Jarvis returns a higher degree of variety. Can I typically address this with parameters or do I need to do something else?

GPT-3 can do translation but remember it was trained mostly on English.

Typically you just do several generations. I would not try and get it to give you multiple samples to choose from for a few reasons:

  1. Token count will be high for every generation
  2. Quality can be lower when you ask for a list

I’m not sure what Jarvis is in this case.