Selfhosting

2 readers

1 users here now

All things selfhosting and homelab related Resources: - https://github.com/awesome-selfhosted/awesome-selfhosted - https://github.com/awesome-foss/awesome-sysadmin

founded 1 year ago

Selfhosted LLM (ChatGPT) - Lemmy.world (lemmy.world)

submitted 1 year ago by operator@kbin.social to c/selfhosted@kbin.social

1 comments fedilink hide all child comments

Stumbled across this on lemmy.world. As we are defederated, reposting here.

I've recently played with the idea of self hosting a LLM. I am aware that it will not reach GPT4 levels, but beeing free from restraining prompts with confidential data is very nice tool for me to have.

Has anyone got experience with this? Any recommendations? I have downloaded the full Reddit dataset so I could retrain the model on this one as selected communities provide immense value and knowledge (hehe this is exactly what reddit, twitter etc. are trying to avoid...)

you are viewing a single comment's thread
view the rest of the comments

[–] walking-octopus@kbin.social 2 points 1 year ago

If you want to host a capable pretrained model, feel free to check out LLaMA, especially the LLaMA.cpp since it allows for speedy inference. For the front-end, there's text-generation-webui, official web UI, Serge, XInference, or chatbot-ui with LocalAI (a server that makes LLaMA.cpp use OpenAI's schema).

For the model fine-tunes, I'd personally recommend WizardLM. It's not perfect, far from it, but it seems the closest to GPT-3.5 in my experience. Be sure to never trust what it says though, it does hallucinate less then other fine-tunes I saw, but still does so frequently enough.

There isn't really much of a need to train a model on a particular community. If you need it to work with changing facts, just throwing results from the search engine into the context window. Most of these models were already trained on huge datasets including Reddit, so...

If you want to fine-tune it on most helpful comments to make sure it generates more consistent advise, I'd recommend QLoRa and a 1k instruction dataset like in LIMA paper. Though again, I'm not sure there's any use for that.