Could you please include the prompt format?
#3
by
Cypherfox
- opened
Greetings,
The prompt format seems to strongly affect the behavior of this; I've had reasonably good luck using ChatML format, but I feel like it's just a little off. I've been using the Q8 and Q6 versions, so I'm going to try the unquantized today and see if ChatML works okay with it.
But generally it's super-valuable to include the prompt format if possible.
Thanks muchly!
-- Cypherfox
Chat template is given in tokenizer_config.json
https://huggingface.co/fblgit/UNA-ThePitbull-21.4B-v2/blob/main/tokenizer_config.json#L32