Marinara's LLM Hub

Join Our Discord Community

Downloads

Note

The newest version is 7.0. These presets are designed for SillyTavern.

Tip

The preset below is for Chat Completion only. For my older Text Completion presets, visit my HuggingFace.

Warning

My presets are designed for chatting, roleplaying, and fiction writing with LLMs. Due to their uncensored nature, they are intended for adults only. By downloading the files below, you confirm that you are 18 years of age or older and acknowledge that I am not responsible for any content you create or share using these presets.

Link	Description
Master Download ᴺᴱᵂ	Downloads the entire .zip file that includes the preset, regexes, logit bias, and character cards
Preset ᴺᴱᵂ	Universal preset alone
Regexes ᴺᴱᵂ	Regexes that customize the prompts' looks, clean formatting errors, remove em dashes, and unify formats
Logit Bias	Logit bias for OpenAI models
Professor Mari	Professor Mari's character card, an assistant who will teach you all about LLMs and help you out with creating character cards
Dottore	Dottore's character card, the Second of the Eleven Fatui Harbingers from Genshin Impact

SillyTavern Spotify Music Extension ᴺᴱᵂ

Want to add some music to your SillyTavern experience? Check out my SillyTavern Spotify Music Extension! This extension allows you to integrate Spotify with SillyTavern for an enhanced chatting experience.

Discord Buddy ᴺᴱᵂ

Hey, while you're here, perhaps you could also check out my Discord Buddy project? It's an LLM-powered Discord bot you can host. Open-source, free, with many, many great features. Check it out!

About

Howdy, I'm Marinara!

An LLM enthusiast, teacher, merger, and prompt engineer since late 2023. Additionally, an artist and coder on occasion. I'm the author of models such as NemoMix Unleashed and many universal prompts, including professional ones for companies like Tolan, MyDreamCompanion, or Toffee.ai. I consult for various AI services and offer commissions for cards, prompts, and artworks.

Here you can find my roleplaying and creative writing presets, as well as recommendations, cards, and guides. All related to LLMs. Feel free to suggest new sections to add or models to test. I hope you'll find it informative, helpful, and fun to read.

Cheers and happy gooning!

Glossary

Term	Definition
LLM	Large Language Model (AI Model)
SillyTavern	Frontend for AI use, the one I create my presets for (https://github.com/SillyTavern/SillyTavern)
Prompt	What you send to the model. This includes everything: instructions, cards, lorebook entries, example messages, and the chat history
Jailbreak	Part of the prompt that is responsible for "jailbreaking," aka "removing censorship" from the model
Preset	A pre-set format of how you send the prompt to the model. These typically involve instructions, optional jailbreaks, and sampler settings
(Character) Card	Typically a description of a character that may include their image, personality, and example message
Persona	Like a character card, but for the character you play as
Scenario	An outline of what happens in the roleplay
Lorebooks (World Info)	Prompt entries that trigger based on activated keywords
Samplers	Algorithms or methods used to decide which token the model should pick next when generating text
Temperature	One of the samplers that controls how creative the output is allowed to be. Numbers closer to zero produce more deterministic outputs, while those closer to 1 and above are more creative
Tokens	Building blocks of text sent to the models. These include symbols, letters, words, and codes
Context	How many tokens the model can process at once, e.g., 32k. The larger the number, the more it remembers
Chat Completion	Format of sending the prompt as a structured conversation, with clear distinctions between different roles (user, assistant, system). Used when you access models via API
User	All the messages that count as being sent from the human (you)
Assistant	All the messages that count as being sent from the AI (model)
System	Instructions for the model, appended before the start of the conversation
Text Completion	Format of sending the prompt to the model as a big wall of text, and asking it to "complete it" (hence the name). Used with locally run models
Input	Your last sent message to the model
Output	Model's response
Logit Bias	A way to manipulate the probability of specific tokens by either banning them completely or encouraging the model to use them more often
Regex	Search language for string, capable of finding specific symbols, words, or phrases, and editing them
Proxy	A way for accessing models via a "middleman"
Function Calling	Models can call actual functions you've defined and programmed, and pass structured data to them
Prefill	Forcing the model to start its output with a specified text. Useful for jailbreaking
Reasoning	Model thinking goes brrr
Refusal	Skill issue

Guides

Video Tutorial

Check out my comprehensive video guide for a visual walkthrough!

Guide To Prompts

Warning

All the toggles on the screenshot above are mandatory, but you may have only one active at a time from their respective sections.

Type decides the overall style of your use case.
- - Game Master: For both group chats and single roleplays, allowing the model to roleplay for all the characters and the narrator.
- - Roleplayer: Specifically for one-on-one roleplays.
- - Writer: For fanfic writing.
Tense decides the tense of the model's writing.
- - Past: Example, "he did it."
- - Present: Example, "he is doing it."
- - Future: Example, "he will do it."
Narration decides the type of narration.
- - Third-Person: Example, "he said."
- - Second-Person: Example, "you said."
- - First-Person: Example, "I said."
POV decides from which point of view the narration will be.
- - Omniscient: POV of a third party, separate observer, who knows what all characters think, perceive, etc.
- - Character's: POV is filtered through what a specific character perceives, thinks, etc.
- - User's: Same as above, but from the user's perspective.

Note

You can juxtapose these into your preferred style. Let's say you want the model to always reply in first person from the respective character's perspective. In that case, you select options "First-Person" and "Character's". If you want a third-person limited narration from your protagonist's POV, you should go for options "Third-Person" and "User's".

Length sets the final length of the bot's response.
- Flexible: You allow the model to choose the response's length dynamically, based on the current scene (short if in a dialogue, longer if the plot progresses).
- Short: Below 150 words.
- Moderate: Between 150 and 300 words.
- Long: Above 300 words.

Warning

My regexes are required for the optional toggles to display properly in the same format as in the screenshot above.

User's Stats tracks your protagonist's statistics and current statuses. These will affect your roleplay.
Info Box shows details about the current scene. Good for maintaining logical continuity.
- Date & Weather
- Time
- Location
- Important Recollections
- Present Characters & Their Observable States
Mind Reading allows you to see the character's thoughts.
Immersive HTML adds active HTML/CSS/JS elements to the narrative.
Randomized Plot Push pushes the narrative forward with a completely random thing. ENABLE ONLY ONCE AND TURN OFF AFTER THAT, UNLESS YOU WANT RANDOM THINGS HAPPENING EVERY TURN.

How To Import A Preset?

Tip

Presets include instructions and settings for models. Different ones offer different experiences. There is no "best" or "worst" preset; it's all a matter of your personal preference. Use the one you like best, and feel free to write your own, custom ones!

Open SillyTavern.
Go to the Chat Completion Presets tab by clicking the first icon from the left (sliders) at the top of the page.
Click the Import preset button (the one with the little arrow pointing at paper) at the top of the screen.
Select the file you want to import.
Open it.

You can access all of your imported presets at any time.

How To Import A Regex?

Tip

Regexes aren't required, but they can improve formatting and fix errors in the outputs. For example, my regexes remove overused em-dashes, fix quotation marks, and turn ... into …, reducing the overall token count.

Open SillyTavern.
Go to the Extensions tab by clicking the third icon from the right (cubes) at the top of the page.
Click on Regex.
Click the Import button (the one with the little arrow pointing at paper).
Select the file you want to import.
Open it.

How To Import A Logit Bias?

Warning

Logit Bias works ONLY with OpenAI models (GPTs). I based mine on Avi's, so all credit goes to them. Logit Bias is a must if you want to use OAI models, however, keep in mind GPT-5 Thinking does not support it.

Tip

Logit Bias allows you to control which tokens are more or less likely to appear in the model's output. The values can be set between -100 (semi-banning the specified token) and 100 (encouraging the model to use a token at every opportunity). Number 0 is neutral. Read more about it here.

Open SillyTavern.
Go to the Chat Completion Presets tab by clicking the first icon from the left (sliders) at the top of the page.
Slide all the way down to the Logit Bias section.
Click the Import preset button (the one with the little arrow pointing at paper).
Select the file you want to import.
Open it.

How To Import A Character Card?

Tip

Character cards are characters you chat or roleplay with. These don't necessarily have to be characters; they can also include assistants, narrators, and even entire worlds.

Open SillyTavern.
Go to the Character Management tab by clicking the first icon from the right (card) at the top of the page.
Click the Import Character From File button (the one with the little arrow pointing at paper).
Select the character you want to import.
Open it.

How To Edit & Toggle Prompts?

Tip

My prompt is plug-and-play, but you may want to edit it or toggle some optional stuff.

Open SillyTavern.
Go to the Chat Completion Presets tab by clicking the first icon from the left (sliders) at the top of the page.
Slide all the way down to the Prompts section.
Click on the small pencil icon on the right side of the prompt to view its contents and edit it.
Click on the small toggle button on the right side of the prompt to enable/disable it.

FAQ

Q: I see some parts of your prompts in {{// this format. }}{{trim}} What does it mean? Won't it confuse the models?

A: This is formatting for comments on SillyTavern. Any text rendered like that won't be sent to the LLM with your request. Only you can read it. The trim part ensures no empty new lines are sent.

Q: Why are some of your prompts empty, but left enabled anyway?

A: Some shared character cards include custom prompts and jailbreaks. They will be placed in those empty prompts.

Q: What preset should I use?

A: Use whatever works for you. If you get too many refusals with mine, try someone else's or write your own. I can't test my prompt with every model in existence, in every possible scenario, with every possible character card.

Q: Can I edit your preset?

A: Of course! I highly encourage it! You can use it as a base to build something personalized, or steal its parts to inject into other prompts. Experiment and find out what works best for you.

Q: Why do you recommend turning off reasoning?

A: Because it's been proven that models do not reason the way we think. Reasoning is obsolete for roleplay/creative writing purposes, and eats up many tokens. It was created with solving complex issues, not gooning in mind, which severely limits creativity and makes the outputs more deterministic. Nowadays, models are smart enough to track the current scene well enough to not have to think about all the details before responding.

Q: What is the best format for writing character cards?

A: Plain text, though, again, everything that works, works. Personally, I've noticed the best results with well-structured descriptions in paragraphs. You can also enclose parts in respective XML tags, such as:

<appearance>
Your character's appearance here.
</appearance>

Q: Can I use AI to generate my cards/lorebooks/presets?

A: Yes, though expect subpar results. The more effort you put into writing these yourself, the better the quality of output you'll get. By using synthetic data, you're encouraging models to fall into established patterns and get 'lazier'. You can always outline your ideas to the model or provide them with an example you want them to follow when generating stuff for you. Except, don't ever use AI-generated prompts; models do not know how to prompt themselves.

Q: Which Prompt Post-Processing do I use?

A: Merge consecutive roles (no tools).

Q: Which frontend do you use?

A: SillyTavern.

Q: What do you have in Advanced Formatting tab on SillyTavern (Text Completion)?

A: Nothing. If you're using Chat Completion Presets, none of the prompts from Advanced Formatting are sent.

Q: I got a refusal?

A: Skill issue.

Recommendations

Recommended Models

Note

Last updated on 03.10.2025

Tip

It's safe to assume the first model on the list is the one I'm currently using for main roleplay. Please keep in mind this is my subjective ranking, based on my own preferences.

Warning

ChatGPT-4o and GPT-4o are two different models! I know, not confusing at all. The one I recommend is chatgpt-4o-latest ('Latte'). The other one, GPT-4o, is pretty bad.

Position	Model	Context	Censorship	Strengths	Weaknesses
1	Sonnet 4.5 ᴺᴱᵂ	1mln	Low	They cooked with this one. It's smart, follows instructions well, has a good writing style, and fixes Opus's repetition issue on higher contexts. I'm hyped for Opus 4.5	You can't really control its style, so if you don't like it, you have to either get used to it or switch a model
2	Claude Opus 4.1	200k	Low	Still the smartest one on the list, with a unique writing style and creativity. Opus can't be beat. It's impossible. The writing is too good. The smarts even more so. Nothing else compares right now. Is TLD still on suicide watch, if Sonnet took the first spot?	Extremely expensive to use regularly and falls into repetetive patterns quickly. Works meh on higher contexts
3	ChatGPT-4o	128k	Low	Best conversational style, big knowledge base, very smart, funny	Unusable without proper regexes and a logit bias
4	GPT-5	400k	Medium	Anthropic won again. GPT-5 falls down a few spots, because it works worse somehow recently? Not sure what they changed, but you can tell it got more robotic than it used to be	A nightmare to prompt, and without a good one, it simply doesn't work. Is absolutely terrible without High thinking
5	Claude Sonnet 4	200k	Low	Perhaps I have treated you too harshly, Sonnet 4. It's still not as creative as its predecessor, but it has its moments	Lacks 3.7's creative spark and is very passive
6	GPT-5 Chat	128k	None	A terrible disappointment at first, but they fixed it, especially with how good the prose feels; now it's a smarter ChatGPT-4o. It's really dumb for me recently.	As all the other OAI models, it requires regexes and a logit bias to be good, plus its prose is worse than ChatGPT-4o's
7	Claude Sonnet 3.7	200k	None	Excellent at both writing and roleplaying, works great on higher contexts (above 32k)	A little outdated
8	Grok 4 Fast (Free) ᴺᴱᵂ	2mln	Low	A very hit-or-miss model - sometimes it's peak, sometimes it produces absolute garbage. It's only higher than Gemini 2.5, because it's available for free as of now	You will have to re-generate a lot with this one
9	Gemini 2.5 Pro/Flash	1mln	Medium	Very smart and good at instruction following. Jumped up in the ranking, since it got better with my newest prompt	Echoing hell ("echoing?"), stiff prose with no logit bias to control some tokens' usage (prepare for a lot of "then" and "didn't"), it's a 50/50 chance the model is lobotomized on the day you use it
10	Kimi K2 0905 ᴺᴱᵂ	260k	Low	Not the smartest, but it's slowly getting there. I have high hopes for these models in the future, given that their creators genuinely care about improving their creative writing capabilities	Still not great at higher contexts (above 32k), and still lacking that extra polish
11	DeepSeek R1/V3	163k	None	Creative and wild	Hard to control and struggles with repetition
12	GPT-4.1	1mln	Low	Good style, doesn't have an annoyingly positive bias, has a large working context window	Unusable without proper regexes and a logit bias, somewhat lacking in smarts
13	Grok 4	256k	None	Grok, is this true?	It's like a teenager going through puberty, so horny and not very smart
14	DeepSeek V3.1	165k	Low	It just works	Worse than V.3 and R1 in every single aspect. What were they thinking
Extra	NemoMix-Unleashed-12B	128k	None	Obligatory shill of my model, some claim it still holds up	Old and small

Warning

I noticed some underlying issues with GPT-5. I'm barely able to control it; it feels like every time I get a generation, it's like a different model responding. Requires further testing. It may be awful, actually. It's just shit. Stick to ChatGPT-4o while you can. They fixed GPT-5 at last, so I've added it to the ranking officially.

Model Providers

Note

If a proxy is listed here, it means I personally tested it and its owner agreed to being mentioned on this page.

Provider	Description
~~Crystal Proxy~~	~~Proxy of my trusted friend. For just $20 a month, you can get access to every big model out there. I highly recommend checking it out, even if it's to hang out with us on her Discord server.~~ No longer accepting new members for now
Anas Proxy	Proxy recommended by a friend. It offers a few models to choose from, including Opus and Sonnet with unlimited contexts. You pay either around $1 for unlimited requests per day of use, or for $2, you can buy 100 requests to use whenever you like. Tested, and everything works
4risu	I accidentally stumbled upon this proxy, but it proved awesome, and the owners are kind. Has a free tier and a paid tier with less downtime and more models to choose from. Just remember to grab an English role if you don't speak Spanish
Mprox ᴺᴱᵂ	Good friend of mine, he's been my main GPT-5 supplier for a while now. He doesn't have as many models to choose from as others, but his endpoints accept parameters such as temperature or logit bias. If you're looking to use OpenAI models, check him out
OpenRouter	I hope this one needs no introduction. This is currently the safest and easiest way to access all the models. Pay-as-you-go when using them

Preset Creators

Creator	Description
AvaniJB	Good prompts, regexes, and logit biases (the last one, I shamelessly stole from them). Also, I love their little cat character
Pixi	The grandparent of prompting. One of my biggest inspirations
Nemo Preset	Personally, I find this preset too big and unnecessarily bloated, but it's still a very well-written and well-thought-out preset! If large token numbers and customization options don't scare you, I suggest checking this one out
Celia	Customizable, creative, overall a great prompt. Plus, Celia is cute
AI Presets ᴺᴱᵂ	Discord server containing many great prompts from amazing creators such as Loggo, Gerodot, Ashu, Chi-Bi, Nara, and others.

SillyTavern Themes

Theme	Description
MoonlitEchoes	It's the one I'm currently using, and the one you can see on my screenshots! It's beautiful, I absolutely recommend it
NemoEngine	Very convenient for prompt management, allows you to play videos and streams as a background for your chats (with sound), and improves on tabs' design.