Good job!
Just wanted to say how GOOD this early tune is already. Gemma 2 is excellent overall, but some of the early tunes seem to lobotomize it to some extent or other. Not Sunfall though - although it's a little difficult to fully pinpoint the benefits in normal conversation, especially when running it with DRY enabled, but isolated one-shot tests show clear differences.
Stock gemma2:
Sunfall:
Thanks for the feedback! Glad you like! It’s a work in progress and hopefully it will only get better as we iterate.
The output of the model is actually more varied than in the original Gemma 2, however, the model has problems when continuing to write a text. (For this purpose, I use "First message" on a specific character card). If the character card is not completely empty, it completely refuses to continue writing. Otherwise it will only write a few sentences. This does not depend on whether the original text is "controversial" or not.
As for "controversial", especially sexual content, Sunfall has no problem with that. Where Gemma 2 creates a verbose refusal, Sunfall writes the desired story.
the model has problems when continuing to write a text. (For this purpose, I use "First message" on a specific character card). If the character card is not completely empty, it completely refuses to continue writing. Otherwise it will only write a few sentences. This does not depend on whether the original text is "controversial" or not.
Weird! I have been primarily testing this on character cards in Silly Tavern. Is there an easy way for me to replicate what you are saying? I'm not sure I understand exactly what the prompt sent to the model looks like when it refuses.
Does this refusal not happen with the Gemma2-9b-it model?