Great model.
Great model with some oddities, first of all I have no idea how to properly use the recommended prompt format as the model outputs robotic and formal responses. Instructions do help a decent bit but it is a night and day difference if you compare it to the roleplay preset in sillytavern. Even furthermore surprisingly it seems to stick to the character less, while roleplay doesn't have that problem.
The latest version is a pretty big improvement, tried 0.60 and wasn't super impressed but this feels quite a lot better already now.
I was very impressed by the prose while using the roleplay preset but it's not useable with this model as it will very often towards the end output nonsense. I'm not sure how to transfer that to the proper prompt format, simply copying the instructions didn't really recreate what I wanted.
A screenshot or preset of your recommended sampler settings would be nice, do we put temperature as last?. Sorry to bother you and thanks for this lovely model, you always come out with the best stuff.
I have no idea how to properly use the recommended prompt format
In SillyTavern, you can import the provided json files with these buttons in the Advanced Formatting tab, then use the model normally like you would with other formats (the prompting format here doesn't use the title, lore and tags roles described in the model card):
A screenshot or preset of your recommended sampler settings would be nice, do we put temperature as last?
I'm still testing it. At temperature = 1 like suggested in the model card, it shouldn't matter whether temperature is last or not. Unfortunately at temperature = 1 there appear to be repetition issues that are difficult to avoid without loss of coherence.
Increasing temperature substantially (up to 2.5-3.0, even), with Min-P around 0.5 and no repetition penalty appears to help with the repetition without making the model incoherent, but it depends on the use case/character.
In SillyTavern, you can import the provided json files with these buttons in the Advanced Formatting tab, then use the model normally like you would with other formats (the prompting format here doesn't use the title, lore and tags roles described in the model card):
I know how to import it, my question was how to use your prompt format effectively, I found that it gave robotic responses, the roleplay preset was a lot better. I think this is the reason.
Is there a way I can apply this to your prompt format and how would I do this?. I'm talking specifically about these little bits of instruction in the output sequence.
"engaging, natural, authentic, descriptive, creative"
I know there is a length modifier so "2 paragraphs" wouldn't be necessary, this shit makes a world of difference.
I'm still testing it. At temperature = 1 like suggested in the model card, it shouldn't matter whether temperature is last or not. Unfortunately at temperature = 1 there appear to be repetition issues that are difficult to avoid without loss of coherence.
Increasing temperature substantially (up to 2.5-3.0, even), with Min-P around 0.5 and no repetition penalty appears to help with the repetition without making the model incoherent, but it depends on the use case/character.
Thank you, I'm struggling to figure out what kind of min p settings I should be using not just for your model but any model. Wish there was a preset that was just objectively the best lol.
Is there a way I can apply this to your prompt format and how would I do this?.
You can use the Mistral-instruct format directly there in the output sequence.
▄
[INST] The following response is: engaging, natural, authentic, descriptive, creative. [/INST]
▀message (length: medium)
The newline at the end is important, don't remove it.
Or you can add a character note at depth 1:
Thank you!!!!, would have never figured this out without your help. I knew that the character note could achieve a similar result but heard somebody saying it's not as good, probably should have tried that also on second thought though. Many thanks for the lightning quick reply.
Great model. Keep up the good work on these!
Just a little feedback, this model doesn't do a whole lot of descriptive writing, often times I will get 2 whole paragraphs of dialogue but barely any description of character's actions, emotions or thoughts. Feels more like a conversation than a roleplay.
I'm using detailed character cards with descriptive first messages so that wouldn't be a problem.
Some of the data is like that, but the majority in terms of tokens is proper roleplay with both dialogues and description (even entire sections of just description of actions and emotions); hard to say where the problem could be, but it's possible that if you only write dialogue in your inputs, the model will respond mostly with dialogue and thoughts.
It's also possible that character notes in the form of [INST] The following response is: engaging, natural, authentic, descriptive, creative. [/INST] (if you've added one as discussed earlier) could lead to more dialogue-like responses, so try without one as well. You should start a new chat if you remove it, since the model will try to follow the format used up to that point.
I have indeed tried it without the character notes to the same effect, I don't ever reply with just dialogues. I'm very well aware that you more or less get back what you put in, I don't have this problem with other roleplay models.
Will see about starting a new chat, for testing purposes I usually just delete everything until my first reply and test the generations. I always assumed new prompt format takes effect immediately?. Either way I'm pretty sure I tried new chats too.
Will see about starting a new chat, for testing purposes I usually just delete everything until my first reply and test the generations. I always assumed new prompt format takes effect immediately?
If you delete everything, it should work in the same way, but other than that I'm not sure how the model can be made more descriptive in its current state. It's probably indeed a bit more dialogue/thoughts-focused than other ones, though.
but other than that I'm not sure how the model can be made more descriptive in its current state. It's probably indeed a bit more dialogue/thoughts-focused than other ones, though.
All good, at the end of the day the model is not even complete yet. I have also found that the character notes don't seem to do much.
Character notes in the form of [INST] ... [/INST] have more of an immediate effect than a long-lasting one and from what I've tested work better with personality/behavior-related instructions rather than general (and somewhat vague) statements like "be engaging, natural, authentic".