Some kind of advice for a v2

#2
by Reality123b - opened

Since you've seen it babble like that, and i've noticed a thing that your model was trained on the whole sample, it wasn't trained on the formatted [INST] tags which would've been formatted like for example, [INST] Hello! [/INST] Hi, how may i help you? -->

Prompt Response
Hello! Hi, how may i help you?
where you'll likely preprocess those tags and make separate instances for separate chats with each turn giving the prompt (or if next turn, previous response+prompt formatted with those inst tags)

Well, thanks for the advice, I kind of didn't think much of that issue. Thanks for letting me know

You're welcome. Always nice to help a fellow builder

Roman190928 changed discussion status to closed

Sign up or log in to comment