bloatmaxx

by meatrocket meatyrocket@proton.me discord: meat.rocket

for Anthropic, OpenAI, Google, Cohere and Mistral LLMs.

Downloads List

ST extension
Latest Updates
Older Presets

commonly asked questions

look forward to new updates, im always improving these presets from time to time

wtf is this?

my RP presets that focus heavily on presentation & interactivity through HTML/markdown, which are also based on the nature of your roleplay
includes a SillyTavern extension to support HTML button inputs from AI (for kino text adventures)

includes toggles for the following:

HTML panels + examples (2.5k + 2k tokens)
Codeblock panels (900 tokens)
Interactive Text Adventure Buttons (500 tokens)

why is it so fucking large? 1k+ tokens is bloat enough

this originally included a bunch of fluff for multiple models to work with.
debloated and bloated presets below (litemaxx vs opusmaxx)
the instruction set alone for the HTML + options excluding any examples is about 2.5k tokens in total.

what is the difference between all your presets?

bloatmaxx is the original version of the 7k token preset intended to be used with GPT-4-1106/Claude 2.1 and newer
opusmaxx has reduced token bloat (~3.5k tokens), and optimized heavily for Claude Opus since the model's release
litemaxx has a much smaller sized JB, as low as 500 tokens (modular)

how to install?

download the desired preset below
import the preset into your sillytavern's chat completion tab
enable prompts/prefill as needed (assistant prefills are at the bottom of the preset)

what model do i use?

whichever is the best, smartest (for RP & complex instruction following) model available for you to use

here is my personal RP ranking between the common models

originally made for furbo. all my presets are tested and confirmed working on the following models:

⎗

Claude 3 Opus / Sonnet
GPT-4-Turbo / GPT-4-32k 
Gemini 1.5 Pro
Mistral Large
Cohere Command-R-Plus

which model do you recommend?

Claude Opus is the best by far with my presets, offering an amazing balance between creativity & intelligence, with instruction-following capabilities that rival or even somtimes beat GPT-4 Turbo. YMMV with Sonnet, however.

GPT-4 Turbo's prose is pretty dry most of the time (also lacking in creativity and initiative somewhat), but it can follow complex instructions well. its massive parameters make it a good model for general logical reasoning

Gemini Pro 1.5 is also a very good model. suffice to say, while it sometimes forget certain details over an extended long context chat, the back and forth dialogue by the model is very engaging and stays true to the character. it also suffers less from the same prose issues as GPT.

how do you get images? how does it work?

enable the prompt and the corresponding prefill toggle (depending on your version of the preset)
it uses an external service ("pollinations.ai") to prompt an image generation via url

retarded meme preset

agreed

downloads

ST extension

Created by horsearmor

clickable html elements in ST, for HTML & text adventure prompts copy https://github.com/horse-armor/st-clickable-inputs into your Extensions manager

Disclaimer: This extension injects a small <instruction> prompt @ depth 0 May result in issues on blank cards or non-HTML based presets. If you don't need it, disable the extension manually.

Latest Updates

[03/05/24]: will maybe look into finding better ways to remove claudisms, for now i am on hiatus, so leave feedback (or shit on it) in the thread (ill see it in the archives), or email me using my burner at the top of this page and i will try and address the issues if & when i find time abroad

normal v1.2 (03/05/24)

v1.2: added a more jailbroken prefill

tweaked instructions a bit, experimenting with new prefill, narrator + fixed some bugs with ST extension working with the preset toggle and choose your preferred settings, some combinations may work better than others. from my own testing, i find that opus likes to talk as you only when it can't find other ways to extend (slowburn) the reply.

started from scratch, contains everything from litemaxx in a different arrangement storyteller for NPCs instead of "you are {{char}}" should work better for scenario/RPG cards

from testing, this one works well for most cards i tried i tried to slim the CoT as much as I can, try it out (usually creates more fleshed out responses from your characters) so far, it should stick to the word limit, limit lewd & generate dense content without speaking as you.

make sure to select your preferred PoV & settings

additional JBs & prompts, with word limiter

regex to remove <CoT> blocks: /[\[\<]CoT[\>\]].*?[\[\<]\/CoT[\>\]]*/ims

regex to remove  tags: //ims

credits to pixijb for inspiring the main prompt. a4a for the anti-lewd (though i got it from someone else)

special thanks to anonaugus for the prefill suggestions

opusmaxx v1.3 (2/05/24)

v1.3: updated instructions + made retarded peanut gallery slop a separate prompt

i don't really recommend this preset anymore but i was asked to make this use this if you preferred the styles from this prompt

litemaxx 1.6 (30/04/24)

v1.6: applied some tweaks to make it work for scenario cards better

i still check the thread archives, so leave your feedback (or criticisms) and ill try to improve upon it special thanks to the anons who can't be credited

Older Presets

normal v1.1 (02/05/24)

v1.1: tweaked instructions a bit, experimenting with new prefill, narrator + fixed some bugs with ST extension working with the preset

normal v1.0 (01/05/24)

v1.0: just a normal preset

as per request, i even tried it with sonnet (without noticing i left it on) in my own slowburn and was decently impressed with the speed and content, even with the CoT included. opus is quite peculiar though so YMMV

lite v1.5 (28/04/24)

v1.5: fixed a bit more of the blatant issues i was seeing with the last preset; halved CoT gen. time (~32 seconds >> 15 seconds with Opus)

returned things back to inject @ 1, likely no longer need CoT to produce consistent results (at least from testing) i still recommend enabling CoT for more higher quality and consistent RPs but i would leave it off for the coom don't forget to enable the instructions you want first...

lite v1.3 (25/04/24)

v1.3: new RP instructions and CoT, tweaked all of the prompts and reduced🫃 bloat HTML prompt rewritten and improved, size decreased from 2400 tokens >> 1250 tokens

added a thinking/CoT prompt, should create cohesion between all instructions much more rigorously

RP focused preset, featuring new style toggles that are completely modular

if you want to add more prompts, insert it somewhere under jb toggles with prompt inject @ depth 0 inspired by many other anons' JBs

lite v1.0 (14/04/24)

(includes toggle for HTML/codeblock panels)

new lightweight jailbreak, using the prompting principles of my other presets as a core function. heavily reduced token usage; 900 token jb + prefill preset for Claude/GPT/Cohere/Gemini focuses heavily on the adherence to character, prose, and initiative for events

opus v1.2 (14/04/24)

v1.2: further refinements, bug fixes & massively trimmed instructions

streamlined things a bit for opus, updated instructions and prose WORKS on claude 3 opus/sonnet, gemini 1.5 pro / 1.0 ultra & gpt-4-32k & gpt-4-1106

includes brand new codeblock panels that are similar in principle to the HTML panels, but smaller in token output. also includes html + text adventure prompts from previous v10

bloat v10 (31/03/24)

updated html instructions, added a vision prompt and a brand new text adventure prompt

(just enable the corresponding prompt + prefill)

(works well with the ST extension)

(generated with sonnet)

special thanks to:

horsearmor (for the extension, proxy owner)
mm (examples, proxy owner)
branon (proxy owner, and for making this card)
lurker (huge help with prompting)
/aicg/ anons & jbmakies for suggestions + ideas
sturdycord