I saw Generative AI for Beginners from Microsoft on GitHub. I’ve looked at https://fmhy.pages.dev/AI but I’m not sure what I’m really looking for.

You know ready, aim, fire? I’m in the AIM stage.

  • CetaceanNeeded@lemmy.world
    0·
    8 months ago

    I’ve been using GPT4All on my laptop and using mostly 7B models due to my RAM limitations and I am amazed how good some of them are.

    It’s been really easy to use. There are models you can download from within the UI or you can get adventurous and download them from elsewhere, they just need to be in the .gguf format. I get most from TheBloke on hugging face.

    So far my favourite has been solar-10.7B-instruct-v1.0-uncensored, it has been astonishingly good.

    • neidu2@feddit.nl
      0·
      8 months ago

      Oooh, do tell me more, please. I’ve been toying with the idea of setting up gpt4all myself, but I haven’t really had the time to look into it very much yet. I have a couple of questions, though:

      • I guess it’s safe to assume that it runs on linux?
      • Is it possible, with some scripting, to provide additional training data, such as connecting it with a wikipedia crawler?
      • By combining it with some script-foo, can I have it also look up stuff for me on the fly, for example "extract THIS kind of information from THAT site ?
      • CetaceanNeeded@lemmy.world
        0·
        8 months ago

        Yes it runs on Linux, my laptop is running Manjaro and I installed it from the AUR. I’m not sure if the scripting is possible, there is an openAI compliant web API you can turn on so maybe possible through that, you would probably have to feed in the content of the site with the prompt though, I’m not sure there is a better way but I guess that sort of behaviour is a bit out of scope for GPT4All.

        There is a local documents feature that allows it to access text files on your machine that you give it specific access to but I think it’s fairly limited in its ability.

      • zeluko@kbin.social
        0·
        8 months ago

        The GPT services out there use something called ‘tools’.
        They get presented to the model and the model can ‘call’ a tool with arguments, which can then extract some data and input it into the context for the model to continue.

        I found out, the models which can run on a normal PC (or even a Laptop) are okay, but not super great. (around or a bit worse than ChatGpt3)
        The good stuff (e.g. Nous-Capybara 31B or the Mistral/Mixtral ones) needs some more memory and compute.