Does this use prompt formatting?
It seems to act for the user frequently (like if it asks something it then responds as the user).. And while using Alpaca format it still does that even though that seems to have been the suggested prompt format for the base model this is based on.
I'm curious about the model's formatting as well, my attempts using either Alpaca or chatml were unsuccessful. The model occasionally fails to generate the stopping token, generating too much text. This is a shame, since the model's output is very-very fresh-looking and doesn't have GPT-isms, if only it could stick to a proper format.
@TheDrummer do you mind sharing SillyTavern templates, including the samplers? Or at least mentioning what format this model was finetuned on?
@mpasila @Araki You can also try the merged versions: https://huggingface.co/TheDrummer/Moistral-11B-v1.1-MERGED-PCT-GGUF
Formatting should be better there. Still very fresh looking. I suggest 5% - 10% for the Moistral experience.