Has anyone successfully used (actually building an app) gpt-pilot with an OPEN SOURCE LLM model?

[phalexo]

I've tried it with GPT-4-turbo-preview and it was making progress, even with some grave errors like modifying working code into junk.

But all my attempts to use local coding models via ollama, simply do not work. They loop, produce bad JSON, and seem to do a lot of useless and potentially bad stuff.

GPT-4 API costs quite a bit of money. I ran up the bill to about $90, just for the UI for login/register screen, and 3 basic UI screens.

Not even touching the backend, except login/register table.

[catarino]

Same here. I've been testing local LLMs (mostly ~7Bs and 13Bs, occasionally ~30B , fine-tuned to code, latest being Llama 3) since about a month ago and everything is a mess.

As @phalexo mentions: They loop, produce bad JSON, and seem to do a lot of useless and potentially bad stuff.

Test with GPT-4 was great.

Obviously, I do not expect 7B local LLMs to have same quality as the big gunned LLMs but judging by some of HF LLM leaderboard stars like StarCoder, DeepSeek, Llama 3 Instruct Coder, etc., I was honestly expecting a very different outcome of this.

Maybe I'm using wrong LLM parameters? I use LM Studio but try not to mess too much with parameters.

[hqnicolas]

@phalexo in my case Ollama-Pilot-CasaOs
I'm Using docker container to run GPT Pilod to Ollama
If you like my Automation Leave a Start ⭐