Error: Stuck after step 5 every time- 'EVERYTHING\_CLEAR' (Back Slash Error?)

[cnaples79]

Hi! I'm on Windows using the latest version of gpt-pilot with OpenRouter as my API.

After answering the initial questions, after step 5 the output reads 'EVERYTHING\_CLEAR', and then stops completely. I've tried re-downloading the repo, used both SQLite and PostgreSQL, and can't seem to get past this. if I hit enter, it says 'No input provided! Please try again.'. If I type something (anything) to continue, it goes into the next step of questions without doing the actual coding/planning.

Every time I try to make a new project, it gets stuck at this step. I've waited 10-15 minutes but nothing happens after the 'EVERYTHING\_CLEAR'. Perhaps some sort of bug/typo as I know it's suppose to say 'EVERYTHING_CLEAR', without the back slash.

I've tried going through more of the questions but I know this isn't correct, as I've used gpt-pilot on a different machine previously and know this isn't how it works. Is this a known bug?

Here are provided pictures:

[cnaples79]

Also, I am using the Mistral-7b-Instruct Model with OpenRouter for my API endpoint. Could this be a reason for the error? It's the only thing I can think of, other than there just being a typo or back slash error in the 'EVERYTHING_CLEAR' command/prompt.

[cnaples79]

I believe I was able to create a temporary workaround: in the pilot/const/llm.py file, for the END_RESPONSE, I changed 'EVERYTHING_CLEAR' to 'EVERYTHING_CLEAR'.

For some reason...when I tell the LLM to say 'EVERYTHING_CLEAR', it actually responds with 'EVERYTHING_CLEAR'. I'm not sure if this is a bug in the code or a bug with the specific LLM I am using (OpenRouter's Mistral-7B-Instruct).

If there is a more permanent solution or fix, please let me know!

Edit: After further testing, this is not a complete workaround. While it does get past the step 5 question and into the Code summary, it gets stuck again at the 'actions' request. It looks like lots of the LLM replies have similar errors, with back slashes right before underscores for some reason... I believe it has something to do with the Mistral-7B-Instruct model needing [INST] and [/INST] added to the prompts. Unfortunately I do not have an OpenAI API I can use to test another working model. If anyone else could test the Mistral-7B-Instruct Model, OpenRouter currently has a 100% discount running so the API is free to use at the moment. I'm curious to know whether or not this is the issue.

[nalbion]

@cnaples79 I don't see the difference between 'EVERYTHING_CLEAR' and 'EVERYTHING_CLEAR'.

I had a bit of a play with Mistral-7b-Instruct last night on fix/rework-product-owner-prompt-for-gpt3. This branch is very experimental at the moment, I was just trying to get through this first step.

https://github.com/Pythagora-io/gpt-pilot/compare/main...fix/rework-product-owner-prompt-for-gpt3

[cnaples79]

@nalbion Were you able to get it past the first step with the Mistral 7B model? Also, I believe it may be the Mistral-7B-Instruct model causing the issue, as the 'Instruct' model requires [INST] and [/INST] when you send it prompts (which gpt-pilot doesn't do).

Sorry, with the formatting it gets rid of the backslash. After step 5 it says 'EVERYTHING\_CLEAR' (with a backslash before the underscore). I'm assuming it has something to do with the model since it appears to be working fine for other users, and I've tried re-cloning / re-initializing the project multiple times with the same result.

[LeonOstrez]

@cnaples79 if underscore is issue then you can try this. Inside /pilot/const/llm.py you can try and change constant END_RESPONSE="EVERYTHINGCLEAR" to something else that doesn't use "", for example just "CLEAR". That way it might work better on other models you try out

[nalbion]

@cnaples79 can you try the branch feature/258-PO-prompt-for-gpt-3.5? It seems to be getting as far as the project summary and technology selection.