I want to train, or more likely fine-tune, a model on about 20 years worth of email and text data that I’ve collected.

The goal would be to train it how to respond like me in simple cases.

It’s there a particular base model I should start with?

I’m also interested in anyone’s experience in doing this kind of thing themselves.

  • ScrawnyStork
    link
    fedilink
    English
    arrow-up
    4
    ·
    1 year ago

    I’m inclined to fine tune/make a Lora for Llama (2? I’m not sure how well it’s supported yet), you definitely won’t want to train from scratch. The big thing seems to be formating the dataset properly, although I haven’t actually done any of this myself yet. this blog seems to have a few good articles about it that you might be interested in reading.