I'm blown away
Such an amazing model, first 7b to actually play my character properly. The internal monologue is simply beautiful, it's crazy how smart this model is, and honestly I think the internal monologue really helps to sell the personality of a character. Truly a legendary model in the making.
I have found that the included instruction and context presets didn't work great for me, the responses were short and dry, but to be fair my character card wasn't following the suggested character description structure and the greeting message was fairly short. On the other hand the roleplay preset worked really well, I got longer and more descriptive responses. It also tended to generate more internal monologues.
'Start Reply With' feature was useful in this case if you wanted to guarantee internal dialogues which is what I used because they were just consistently really good. If anyone wants this just place a single asterisk or underscore.
Thanks for your hard work, I haven't seen any model like this before, you got something special here.
EDIT: Edited for clarity.
Thanks; after a few more training samples (43 out of 100 planned) the model appears to be somewhat better than this one but I haven't updated it yet here and I'm not sure if I should create a new repo (and hide the old one) or just modify the files and description here.
But what do you mean exactly that the character card wasn't following the recommended structure? It's meant to be used with the provided SillyTavern context settings.
I also didn't add an internal monologue in the greeting with the provided character card because I wanted to make it clearer that the character may spontaneously come up with them, but perhaps I could update it with one that contains them.
Thanks; after a few more training samples (43 out of 100 planned) the model appears to be somewhat better than this one but I haven't updated it yet here and I'm not sure if I should create a new repo (and hide the old one) or just modify the files and description here.
Updating the repository should be fine.
But what do you mean exactly that the character card wasn't following the recommended structure? It's meant to be used with the provided SillyTavern context settings.
Sorry I worded that badly, it wasn't your character card that I was using but my own and the character description wasn't formatted in a manner like your character card Charlotte.
I also didn't add an internal monologue in the greeting with the provided character card because I wanted to make it clearer that the character may spontaneously come up with them, but perhaps I could update it with one that contains them.
Yeah I completely understand, I actually found that the frequency of internal monologues was much higher when using the roleplay preset as opposed to the Limamono instruct and context presets included. With my own character card I decided to use the 'Start Reply With' feature to guarantee internal monologues as the character was two faced and manipulative which resulted in some really fun internal monologues.
Yeah I completely understand, I actually found that the frequency of internal monologues was much higher when using the roleplay preset as opposed to the Limamono instruct and context presets included.
After extensive testing with newer updates (of which I uploaded the latest earlier today) over the past several days, I think using (length = long) is acceptable (unlike what I initially wrote, where I suggested not to use a length modifier for bot messages) and gives a good frequency of internal thoughts/monologues. So, you might want to try with that also.
I've updated the model card, as well the presets, to reflect this change.
I haven't tried the "start reply with feature"; often monologues can occur also in the middle or the end of the messages, not necessarily always at the beginning. Anyway, I don't think it should be a problem to use that if it works for you.
After extensive testing with newer updates (of which I uploaded the latest earlier today) over the past several days, I think using
(length = long)is acceptable (unlike what I initially wrote, where I suggested not to use a length modifier for bot messages) and gives a good frequency of internal thoughts/monologues. So, you might want to try with that also.
Will give it a beep, thank you.
I haven't tried the "start reply with feature"; often monologues can occur also in the middle or the end of the messages, not necessarily always at the beginning. Anyway, I don't think it should be a problem to use that if it works for you.
Everyone likes their roleplay different. Another feature that I really liked was how sounds are basically verbalized and described. This is the only 7b model that does it properly. Very immersive stuff. This is my favourite model and the best 7b for roleplay as far as I'm concerned.
Also please, Q5_K_M is ideal for us GPU poor, if you're able please upload.
Just a heads up, you got the instruct and context presets mixed up.
I can upload a Q5_K_M version—uploading right now actually.
Thanks for reporting that the instruct and context presets were mixed up, I'll update them once the Q5_K_M model is done.
I can upload a Q5_K_M version—uploading right now actually.
Thank you, you rock.