sphiratrioth666/SillyTavern-Presets-Sphiratrioth

Simple, effective, universal presets for Silly Tavern role-playing purposes. Organized file names, organized preset names, detailed instruction and explanations.

SillyTavern (Version): 1.12.6+
Templates (Included): Mistral, ChatML, LLAMA3, Alpaca, Metharme/Pygmalion
Models (Suggested): a) Mistral Nemo - Instruct (12B); b) Mistral Small - Instruct (22B); c) Magnum V2/V.3 (12B/34B); d) Rocinante (12B); e) Nemo Unleashed by Marinara (12B); f) ArliRPG Models (12B/22B); g) Theia (21B); h) Cydonia (22B); i) Celeste 1.5/1.9 (8B/12B); j) Stheno 3.2 (8B), k) Silicon/Loyal Macaroni/Kunoichi Maids (7B), l) any other models using those exact templates

> 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].

Don't be the Amazon's Saur-off. Be a true Lord of the Templates.

Creative Commons Attribution-NonCommercial-ShareAlike 3.0 License (https://www.deviantart.com/selrond/art/One-Ring-To-Rule-Them-All-507183083)

FILES (.JSON) INCLUDED:

CONTEXT TEMPLATES (STORY STRING)

INSTRUCT TEMPLATES

SYSTEM PROMPTS

SETTINGS (SAMPLERS)

BONUS/OPTIONAL: REGEX

SAMPLERS (USED):

Min-P

Temperature

DRY (prevents repetition)

> 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 VARIANTS:

Sphiratrioth [T=0.7] (Temperature = 0.7) - less creative responses, more precise, tames down the "fidgeting" models

Sphiratrioth [T=1.0] (Temperature = 1.0) - standard, balanced responses, works well with a majority of modern models

Sphiratrioth [T=0.7] (Temperature = 1.3) - creative boost - but without going off the rails, which starts around T=1.5

MESSAGES FORMATS:

1st Person / 3rd Person (switch through System Prompt Presets!)

DIALOGUES - quotation marks if you do not use my REGEX/plain text, no quotation marks, if you add my REGEX

NARRATION - italics

INNER THOUGHTS - code block aka ```Thoughts```

Trim Incomplete Sentences: ON (it prevents messeges from being cut in half with short maximum tokens setting under "conversation" mode preset, you need to turn it on manually in the newest version or...); 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. 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. You can also turn quotations ON/OFF this way - since LLM will generate with quotations now.

PRESETS:

A) Sphiratrioth (Conversation) - minimal narration - when you want to simulate a natural, human-like conversation with a bot. For those who do not like the "verbose" role-play. <<< conversation-like, ~ 150 tokens responses, mostly dialogues with basic narration when needed, it works well >>>

B) Sphiratrioth (Roleplay) - balanced RP - when you want to experience a balanced role-play with responses around 1-2 paragraphs. <<< 1 short paragraph, ~ 250 tokens responses, around 1/3 dialogues against 2/3 narration >>>

C) Sphiratrioth (Story) - AI becomes your Game Master (GM) - when you want those long, verbose paragraphs with rich narration & story-telling. <<< 1-3 paragraphs, ~ 350 tokens responses, around 50/50 narration against dialogues, pushes the story on its own >>>

D) 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 >>>

(Yeah, I was sleepy - {{char}} wanted lush greenery but it accepts it's Autumn, I actually did not read properly what LLM came up with in the first place, fault on my side :-D But you get the idea of a preset at least, haha.)

How to use:

Advanced Users: Enjoy, modify, forget.
Beginners: Read further.

1. 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 >>>
2. 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.>>>/
3. Download my presets.
4. Copy the [Sphiratrioth-(...)-CONTEXT TEMPLATE] files to (...)\SillyTavern\data\default-user\context.
5. Copy the [Sphiratrioth-(...)-INSTRUCT TEMPLATE] files to (...)\SillyTavern\data\default-user\instruct.
6. Copy the [Sphiratrioth-(...)-SETTINGS] files to SillyTavern\data\default-user\TextGen Settings.
7. Copy the [Sphiratrioth-(...)-PROMPT] files to (...)\SillyTavern\data\default-user\sysprompt.
8. Open up Silly Tavern UI.
9. Click on the "Capital A" tab in Silly Tavern UI (AI Response Formatting).
10. Load up my Context Template (Story String) Preset from the Context Templates list.
11. Click on the "Enable Instruct Mode" button (ON/OFF next to the name "Instruct Template").
12. Load up my Instruct Template Preset from the Instruct Templates list.
13. Load up my System Prompt from the System Prompts list.
14. Click on the "Sliders" tab in Silly Tavern UI (AI Response Configuration).
15. Load up my Settings Preset from the Text Completion Presets list.
16. Switch between the 1st Person/3rd Person narration by switching the System Prompt Presets. Theoretically, it should re-evaluate the prompt but it may not work unless you start a new chat!
17. Switch between Conversation/Roleplay/Story Modes & Creativity (Temperature) by switching the Settings (Text Completion Presets/Samplers). Add optional Regex if you want.
18. 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.**

REGEX:

1. Download the REGEX files.
2. Open up SillyTavern Extensions tab.
3. Find REGEX.
4. Import downloaded files as GLOBAL.
5. Adjust order: starts with "trim incomplete", followed by: "find last special character", followed by: "remove **"", followed by "remove ``````"", followed by: "delete quotes" (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.
   
   

PLANS FOR THE FUTURE:

Updating the presets pack to keep up with the new, better models coming out in the future.

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 = 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 :-)

[1] : Virt-io, Silly Tavern Presets, 2024

[2] : Marinara, Silly Tavern Presets, 2024