Best Memory/World Info/Author's Note format for Arcania?
Been playing with this for a few days, tweaking parameters and still loving it as a lighter alternative for Astoria-70B
I'm wondering if you have any guidance on what the best formatting for memory/world info/author's note would be for this model, based on it's training set?
Square-bracket based formats seem to work ok for Memory/WI, but if I put those in AN it 'leaks' into the responses. (Example: [Author's Note: Write a description of the current location.] would sometimes return a result that ended with a 'new' author's note, or include random bits of text in square brackets, or vomit a bunch of square brackets at the end of the response.)
Changing params mitigates some of these issue, but doesn't fully get rid of them. All-in-all some minor bracket vomit is acceptable for the overall quality I'm getting.
Been playing with this for a few days, tweaking parameters and still loving it as a lighter alternative for Astoria-70B
I'm wondering if you have any guidance on what the best formatting for memory/world info/author's note would be for this model, based on it's training set?Square-bracket based formats seem to work ok for Memory/WI, but if I put those in AN it 'leaks' into the responses. (Example: [Author's Note: Write a description of the current location.] would sometimes return a result that ended with a 'new' author's note, or include random bits of text in square brackets, or vomit a bunch of square brackets at the end of the response.)
Changing params mitigates some of these issue, but doesn't fully get rid of them. All-in-all some minor bracket vomit is acceptable for the overall quality I'm getting.
MoEs can be a little unhinged at times especially with prompt types. I've noticed Arcania has the tendency to repeat at times as well. I think this is going to be have to addressed in the next model but if I find a workaround I'll let you know!