Here, you can buy me a Coffee. Just click on the banner above. All the work I do remains free - but I drink a lot of coffee, actually, haha - so it is very nice when you show support by fueling my caffeine addiction :-D
As a bonus, you can also find a lot of my other, side (hobby) projects. Some Reshade presets for different games, my custom TTRPG system (soon), maybe other goodies in the future.
> All the other samplers neutralized. That's all you need in modern times (2024/2025). Samplers order is also correct - so min-p before temperature. > I am not a fan of XTC, it feels too "loose".
CREATIVITY:
Sphiratrioth - Grounded (Temperature = 0.7 or 1.0 or 1.3) - instructions following: PERFECT, formatting: PERFECT, creativity: MEDIUM, repetition: LOW
Sphiratrioth - Creative (Temperature = 0.7 or 1.0 or 1.3) - instructions following: GOOD, formatting: MEDIUM, creativity: HIGH, repetition: LOW
Sphiratrioth - Classic (Temperature = 0.7 or 1.0 or 1.3) - instructions following: PERFECT, formatting: PERFECT, creativity: MEDIUM, repetition: MEDIUM
RESPONSE LENGTH:
Sphiratrioth - [150T] (Max Tokens = 150) - when you want a more conversational feeling
Sphiratrioth - [350] (Max Tokens = 350) - when you want a more roleplay/storytelling feeling
FORMATTING:
1st Person / 3rd Person (switch through System Prompt Presets!)
Trim Incomplete Sentences: OFF - you can use my REGEX files instead, which are better than a basic SillyTavern tool in trimming the unfinished sentences. Additionally, doing it through REGEX allows trimming even the extremely short messages and controlling if you want the * or " markdown signs in your roleplay or not. You can set the output toxens to crazy values such as 20-50 and it will trim everything perfectly, which is not doable otherwise - only through proper regex. The endless issue with too long messages finally solved! Look for Regex files is a separate folder to download now - you import them under SillyTavern Extensions Tab -> REGEX, in order provided in the files readme.
PRESETS:
A) Sphiratrioth (Grounded/Creative/Classic) [150T]<<< 1 short paragraph, ~ 150 tokens responses, around 1/2 dialogues against 1/2 narration >>>
image/png
B) Sphiratrioth (Grounded/Creative/Classic) [350T]<<< 1-3 short paragraphs/1 long paragraph, ~ 350 tokens responses, around 1/3 dialogues against 2/3 narration >>>
image/png
C) Sphiratrioth (GM) - you are a Game Master (GM), AI becomes your player - tabletop RPG-like experience from a GM-POV!
<<< 1-3 paragraphs, ~ between 150-350 tokens responses, around 50/50 narration and inner thoughts, capable of dialogues when situation calls, follows your story >>>
image/png
(Yeah, {{char}} wanted lush greenery but I actually did not read properly what LLM came up with in the first place, fault on my side :-D Still, you get the idea of a preset at least, haha.)
Find your model's native INSTRUCT TEMPLATE. Every fine-tune stands on a base model, which is usually specified by its creator in a manner of "We're presenting the first fine-tune of Mistral Small" or "Nemo on steroids" or "LLAMA 3 with our special sauce" etc. It may be confusing, it may be clear. Not my fault :-P However, you need to identify the base model of a fine-tune by reading its description page. Different templates may be used by creators (tuners) for training - and then - templates suggested on a fine-tune's page work best. However, creators often just choose the basic template of a model they're tuning - so make sure to check the model's description page or assume it's using the base model's template (template of the main model that a given tuner is tuning).
<<< Most typical templates: Mistral, ChatML, LLAMA3, Alpaca, Metharme/Pygmalion >>>
Golden Rule: Match CONTEXT TEMPLATE (Story String) and INSTRUCT TEMPLATE with your model. For instance, when Mistral-Nemo fine-tune is using the Mistral Instruct Template (as specified by creator of a fine-tune) - pick up the "Sphiratrioth - Mistral [CONTEXT TEMPLATE]" file and "Sphiratrioth - Mistral [INSTRUCT TEMPLATE]" file. It's very easy. I keep my naming clear - both the file names and preset names inside of the Silly Tavern UI list when you load them up. They're always clear and distinguishable.
<<< This rule may be broken when you become the advanced user. Then, you'll know what you're doing so you'll mix different context/instruct templates. For now - one ring... khem - one template to rule them all. Don't be the Amazon Saur-off. Be a true Lord of the Templates.>>>/
Download my presets.
Copy the files from [context] to (...)\SillyTavern\data\default-user\context.
Copy the files from [instruct] to (...)\SillyTavern\data\default-user\instruct.
Copy the files from [TextGen Settings] subdirectories (choose the JSON files you want, do not copy the sub-folders!) to SillyTavern\data\default-user\TextGen Settings.
Copy the files from [sysprompt] to (...)\SillyTavern\data\default-user\sysprompt.
Open up Silly Tavern UI.
Click on the "Capital A" tab in Silly Tavern UI (AI Response Formatting).
Load up my Context Template (Story String) Preset from the Context Templates list.
Click on the "Enable Instruct Mode" button (ON/OFF - next to the name "Instruct Template").
Load up my Instruct Template Preset from the Instruct Templates list.
Click on the "Bind to Context" button (ON/OFF - next to the name "Instruct Template").
Load up my System Prompt from the System Prompts list.
Click on the "Sliders" tab in Silly Tavern UI (AI Response Configuration).
Load up my Settings Preset from the Text Completion Presets list.
Switch between the 1st Person/3rd Person narration & the universal or [SX-4] experience by switching the System Prompt Presets. Theoretically, it should re-evaluate the prompt but it may not work unless you start a new chat!
Switch between Grounded/Creative/Classic Modes & Response Length by switching the Settings (Text Completion Presets/Samplers).
Add my optional Regex if you want (read further for more instructions).
Enjoy.
BEWARE:
**IF YOU WANT THE PRESETS TO WORK PROPERLY, REMEMBER TO START A NEW CHAT FOR TESTING PURPOSES/ACTUAL ROLEPLAY. IF YOU "HOT" SWAP THE PRESETS, YOU MAY EASILY FALL INTO A TRAP OF CONTINUING THE PREVIOUS PROMPT. IN RESULT, THE SHORT/BALANCED/LONG RESPONSES AND NARRATION IN 1st/3rd PERSON WILL NOT SWITCH PROPERLY. ALWAYS START A NEW CHAT. THEORETICALLY, SWITCHING A SYSTEM PROMPT FORCES THE PROMPT EVALUATION BUT IN REALITY - IT OFTEN BREAKS, IT DOES NOT WORK PROPERLY. IT WORKS WHEN YOU START A NEW CHAT THOUGH - OR WHEN YOU RELOAD THE WHOLE SILLY TAVERN UI.**
SX-4/GM-4/CG-4:
Everyone: Everything is the same except of the system prompts. SX-4/GM-4/CG-4 roleplay environments (my custom work, check my other posts) differ only through sys prompts. You can just use the universal Context Template, Instruct Template, Samplers and REGEX but with SX-4 format sys prompts. Copy-paste them just like universal sys-prompts, load them up in Silly Tavern just like you do with universal sys prompts. It's that simple.
BEWARE:
> Match CONTEXT TEMPLATE & INSTRUCT TEMPLATE (by names - while loading them up in Silly Tavern UI) ["CAPITAL A" - SETTINGS TAB]. > Switch 1st/3rd Person LLM Narration through SYSTEM PROMPT (start new chat to be sure that it works) ["CAPITAL A" - SETTINGS TAB]. > Switch CONVERSATION/ROLEPLAY/STORY LLM responses style & MODEL'S CREATIVITY through SAMPLERS ["SLIDERS" - SETTINGS TAB].
REGEX:
Download the REGEX files.
Open up SillyTavern Extensions tab.
Find REGEX.
Import downloaded files as GLOBAL.
Adjust order: starts with "trim incomplete", followed by: "find last special character", followed by: "remove **"", followed by "remove ``````"", followed by: "no italics" (optional) and "no quotations" (optional)
Turn Smooth Streaming in SillyTavern Settings OFF. REGEX will make your streaming "laggy" since it corrects the messages as they're generated live but - it's still so much better than unfinished sentences with lack of closing * or " or ``` signs. You can leave the normal streaming on, it's ok - just turn the Smooth Streaming OFF. It is the only way to trim very short messages around 20-50 tokens properly - if that's what you want - and it generally does its job much better than SillyTavern on itself with all normal message lengths. It's worth it.
Troubleshooting: if it does not allow you entering any special sign at the end of your message - open up the regex definitions (in SillyTavern, just edit them) - and uncheck the box for user messages. It will only affect the AI output then.
REQUEST A PRESET:
Please, do not do that. If something is not here, it means that I am not using a model you want and I cannot adjust my presets properly to work with it. I do not want to give you a bad quality product that has not been tested properly, you know. Typical tunes in recent times utilize the provided formats. Qwen uses ChatML, Gemma ones use ChatML or Gemma2 standard SillyTavern Context/Instruct templates. I do not work with APIs for roleplaying and I cannot run anything above 70B locally so forgive me, folks - but please, do not request more presets than provided here :-D
ANNOTATIONS:
I am using the full names of what's what but people simplify and people are great but messy - thus - you'll come across the words such as context or instruct meaning completely different things as you're continuing your journey in the LLMs community. No one writes context template vs context length or instruct template vs instruct mode/instruct model. Everyone simply uses context & instruct, also a story string (since it's clear it is a part of the context template, isn't it? Isn't?! Of course it is! Everyone knows at birth!). It all causes confusion for beginners. Here - a small guide for you - to understand how others operate.
Context = Context Template or Story String (IN TERMS OF PRESETS)
I'm simplifying here but people often get confused with names. Story String is a part of the Context Template - to be precise - but creators of models and presets use those terms as equivalents...
Context = Context Length of a model (IN TERMS OF PROPER USE OF A WORD). People simplify and use mental leaps. Not my fault, again :-P
Instruct = Instruct Template (IN TERMS OF PROPER USE OF A WORD). Again, people simplify...
Model Name - Instruct Model = a model trained to follow instructions in a given instruct template (IN TERMS OF PROPER USE OF A WORD). You catch the drill by now, don't you? :-P
There are other great presets from Virt-io or Marinara available on the site but they come with messy preset names, less clear file structures or forced instructions such as including the character example messages through instruct/context template, which becomes the unnecessary bother. Also, they do a bit less than mine.
However, both Marinara and Virt-AI are great creators and great people - so when you learn and understand how presets actually work - be sure to give their collections a try! You may like the more, you may prefer this, who knows :-)
