Suggested samplers?
Hey there! First thing first, thanks for your new model! I'm trying it with the same config I used with Cydonia 4.3 and Magidonia 4.3 (both really great) but this one takes the narrative one step forward. Even using the Q4_K_S quantization, the "awareness" and "richness" of the generated text relative to the context feels so amazing that it became my new fav model even if I have to "shrink" the context size to 24k to fit it in my VGA. Not that matters too much bc 24b models use to show some degradation between 15 and 25k.
It can handle several chars and retain their characteristics and features, and even allows them to "evolve" respect to the initial features descripted in the char card.
One of the things I perceived is that, around 15-18k context, the generated texts starts to mix stuff. Kinda actions that the char supposedly should do are done by the user and vice-verse. Nothing that a re-roll can't fix.
My current settings:
Mistral v7 Tekken
Temp=0.8
TopK=40
TopP=1
MinP=0.075
Hey there! First thing first, thanks for your new model! I'm trying it with the same config I used with Cydonia 4.3 and Magidonia 4.3 (both really great) but this one takes the narrative one step forward. Even using the Q4_K_S quantization, the "awareness" and "richness" of the generated text relative to the context feels so amazing that it became my new fav model even if I have to "shrink" the context size to 24k to fit it in my VGA. Not that matters too much bc 24b models use to show some degradation between 15 and 25k.
It can handle several chars and retain their characteristics and features, and even allows them to "evolve" respect to the initial features descripted in the char card.
One of the things I perceived is that, around 15-18k context, the generated texts starts to mix stuff. Kinda actions that the char supposedly should do are done by the user and vice-verse. Nothing that a re-roll can't fix.
My current settings:
Mistral v7 Tekken
Temp=0.8
TopK=40
TopP=1
MinP=0.075
Im very new to SillyTavern. Where do i get the Mistral V7 Tekken settings from? Ty for any help!
Hey there! First thing first, thanks for your new model! I'm trying it with the same config I used with Cydonia 4.3 and Magidonia 4.3 (both really great) but this one takes the narrative one step forward. Even using the Q4_K_S quantization, the "awareness" and "richness" of the generated text relative to the context feels so amazing that it became my new fav model even if I have to "shrink" the context size to 24k to fit it in my VGA. Not that matters too much bc 24b models use to show some degradation between 15 and 25k.
It can handle several chars and retain their characteristics and features, and even allows them to "evolve" respect to the initial features descripted in the char card.
One of the things I perceived is that, around 15-18k context, the generated texts starts to mix stuff. Kinda actions that the char supposedly should do are done by the user and vice-verse. Nothing that a re-roll can't fix.
My current settings:
Mistral v7 Tekken
Temp=0.8
TopK=40
TopP=1
MinP=0.075Im very new to SillyTavern. Where do i get the Mistral V7 Tekken settings from? Ty for any help!
Hey! AFAIK, Mistral V7 Tekken context and instruction templates are included out of the box for Silly Tavern!
Hey there! First thing first, thanks for your new model! I'm trying it with the same config I used with Cydonia 4.3 and Magidonia 4.3 (both really great) but this one takes the narrative one step forward. Even using the Q4_K_S quantization, the "awareness" and "richness" of the generated text relative to the context feels so amazing that it became my new fav model even if I have to "shrink" the context size to 24k to fit it in my VGA. Not that matters too much bc 24b models use to show some degradation between 15 and 25k.
It can handle several chars and retain their characteristics and features, and even allows them to "evolve" respect to the initial features descripted in the char card.
One of the things I perceived is that, around 15-18k context, the generated texts starts to mix stuff. Kinda actions that the char supposedly should do are done by the user and vice-verse. Nothing that a re-roll can't fix.
My current settings:
Mistral v7 Tekken
Temp=0.8
TopK=40
TopP=1
MinP=0.075Im very new to SillyTavern. Where do i get the Mistral V7 Tekken settings from? Ty for any help!
Hey! AFAIK, Mistral V7 Tekken context and instruction templates are included out of the box for Silly Tavern!
Found it thank you! I know it might be weird to ask here about off topics. Do you know how to properly use Qwen3.6 to work? I cant see any Templates for this model. Ty if you know!
Hey there! First thing first, thanks for your new model! I'm trying it with the same config I used with Cydonia 4.3 and Magidonia 4.3 (both really great) but this one takes the narrative one step forward. Even using the Q4_K_S quantization, the "awareness" and "richness" of the generated text relative to the context feels so amazing that it became my new fav model even if I have to "shrink" the context size to 24k to fit it in my VGA. Not that matters too much bc 24b models use to show some degradation between 15 and 25k.
It can handle several chars and retain their characteristics and features, and even allows them to "evolve" respect to the initial features descripted in the char card.
One of the things I perceived is that, around 15-18k context, the generated texts starts to mix stuff. Kinda actions that the char supposedly should do are done by the user and vice-verse. Nothing that a re-roll can't fix.
My current settings:
Mistral v7 Tekken
Temp=0.8
TopK=40
TopP=1
MinP=0.075Im very new to SillyTavern. Where do i get the Mistral V7 Tekken settings from? Ty for any help!
Hey! AFAIK, Mistral V7 Tekken context and instruction templates are included out of the box for Silly Tavern!
Found it thank you! I know it might be weird to ask here about off topics. Do you know how to properly use Qwen3.6 to work? I cant see any Templates for this model. Ty if you know!
I haven't used 3.6. I remember I used some 3.5 finetune models and those included a link with suggested context and instruction templates.