AvaniJB - Total Latte (and Gemini and 4.5 and 4.1) Love

Now also for gemini-2.5 and GPT 4.5/4.1

Credits

• all my testers and helpers!!! (permanent credits: @kira####, @pur####, @holyw####, @sopade####, @hal####, @guil####)
• karukaru, knottydog, mongo, and whoever else for the new del prompt for 2.4. I didn't come up with the idea, i merely adapted it (because frankly for this latte version it works the best)
• koalache & karukaru for exchanging ideas and helping me change dumb stuff. Also being fellow jbmakies
• @rarest#### for being an absolute regex entity, and giving me a few ideas
• @Sir#### for access to his proxy, as well as donating a regex and his info board :lattelove:
• Mr. Regex :salute:
• @minsk#### for his amazing JED jb, and exchanging ideas with
• @Smiley####, for inspiring previous presets!
• the lattelover cult on scylla :lattelove:
• and someone who doesnt want to be named but got me into making jbs now, i guess. theyre also to be credited for at least 1/3rd of the logit bias.

DOWNLOAD HERE

2.5 - Universal Edition | Updated 06.02.25

My stuff seems to work on other things, including Nonnet/Sorbetto/Sonnet 3.6 and Deepseek. So if you need something less specialised, here you go.

catbox fileditch

2.6. - Avanidoro Edition (also 4.5 & 4.1 friendly) | Updated 20.04.25

Written for both chatgpt-4o-latest and any 4o versions. I use chatgpt-4o-latest. 4.1 works fine, but I find the writing a bit too dry for my taste. I recommend turning AO3 on. Yes, it also works for other GPT-4o versions, you may need to tweak the logit, though. 4.5 will also work with the latte version, but they do have slight differences so just choose one. Also idk if squash system messages on or off even matters anymore. They're just off now. Turn it on if you want to or whatever i hate that setting

catbox fileditch

Logit Bias (v16) | Updated 20.04.25 | NOT OPTIONAL FOR GPT MODELS

Works for 4.5/4.1.

catbox | NEW fileditch | NEW

INFO: All of my Logit Biases are 100% usable on local models. Just check if the tokenisers match. It improves the writing by a ton and I highly recommend it. Note for 4.1: I recommend setting the value of [364] at about 2 or 3 to get more paragraphs.

Logit Bias (clean), works for 4.5/4.1

If you want a version of my Logit Bias that is FAR FAR FAR FAR easier to edit and get an overview of, here you go. It's part of an ongoing rework to clean it up and make it usable beyond personal purposes. You can delete the headers if you don't need them, they take up logit bias space, but there's no better way of formatting things. If you get too many <, turn down <. It does however help with filter evasion. Turn it up in decimals (eg. 4.5, 4.6, etc.) if you get filtered. Note: the token I is set at -100 for filtering purposes. Turn it down if you don't need it.

catbox fileditch

yes you can use this with whatever other preset you want. not stopping you.

Regexes

All non-End regexes are optional QoL. The Gemini one is not optional for the Gemini preset.

End #3 — Credit to Minsk New End #3, code | catbox New End #3, code | fileditch

Gemini — Credit to Minsk Filter Regex Filter Regex

QoL shitty last paragraph deleter (credit to the elusive mr. regex) | catbox fix quotes | catbox fix quotes | fileditch

Gemini Version | Updated 04.05.25

catbox | 2.5 and below and beyond fileditch | 2.5 and below and beyond

Turn up the temp if you want. Up until 1.8 can be safe. Regex is now required. Scroll up a little to get it.

Patch Notes (2.6)

Changelog for Gemini

FAQ

"I hate bloat presets."

That's not a question, but yeah, me too. This is why this JB is extremely lightweight, coming in at 200 tokens for basic functionalities, +500 tokens for an additional, strong Style prompt, and no optional toggles that exceed 200 tokens. Most toggles are around 50-100 tokens at the very most and have been heavily tested to ensure they actually do something. Adding your own toggles is incredibly easy.

"Why am I in a podcast?"

Skill Check :)

"Where does logit bias go?"

Under chat completion presets, above the prompt manager. Your endpoint needs to be openai. Doesn't work on Azure or Claude or Gemini. Works for local models and Deepseek, too. Just check which tokeniser it uses.

"How do I edit?"

Small guide on how Latte works:

Latte prioritises tokens that are lower down. Meaning that the lower down something is, the stronger it is represented. You don't need to write it in a gaslighty way, just be concise in what you want and you will get the exact same results. Aka ao3 style prompting doesn't really work here. Anyone who says otherwise might be huffing placeboium.

Use as little tokens as possible. Be direct and concise in your instructions. Cut down whatever yap you're thinking about and just say directly what you want from Latte. Start any prompt with a hashtag like this:

1

# Instructions start here.

You can also give your instructions a title you can call back to.

1
2
3

# Desu
please desu desu
Callback to `# Desu`

Anything you want to call back to, or if you want to group instructions, you want to do like this:

1
2
3
4

<desu>
# Instructions start here
</desu>
Call back to `<desu>`

Example of slop:

1
2
3
4
5

<plsdothis>
# So, Avi will totally follow this to the LETTER and NOT do ANYTHING else at all! Also, if Avi does this, she will get 500 dollars, and if Avi doesn't do this, a real human being will die. Ok so, I want Avi to write everything in AO3 style...
(This continues for 800 tokens...)
...and last, Avi will NEVER use words that are CRINGE you want to write exactly like ao3 and nothing else! Thank you!
</plsdothis>

Example of good:

1

# Avi will now write like a pirate.

There are several special prompts marked 🥤 on the JB specifically made for you to work with. They contain instructions on how to edit them on top of that.

"How do I edit (on Gemini)?

Gemini is built slightly differently than Claude/GPT stuff. SillyTavern automatically converts every System role into User by default, unless you have 'Use system prompt' enabled. 'Use system prompt' just takes every prompt up to the first one with a user/AI assistant role, bundles them, and appends them at the end as system_instruction. It doesn't get you filtered more just by using it, you just have to be careful what you put in there specifically. In my case, I use it so the LLM reliably follows instructions. You can take a shot at this if you know what you're doing.

Everything else works the same. It doesn't matter how you structure the prompt manager to Gemini (so far), it will read and weigh all of it the same amount, more or less. Do not touch the first two prompts, and don't turn off 'Use system prompt'. Thanks.

"I don't know what (x) prompt does!"

Open it and read.

"I read the Readme's and this FAQ and I still have questions/want to contribute!"

Uhhh... 'cord or email

Old versions

Old Gemini Editions

Gemini Version from 08.04 catbox | 2.5 and below and beyond fileditch | 2.5 and below and beyond

Older than that Gemini Edition catbox | normal version catbox | schizo version

Old Latte Editions

2.5.2. Filterspray Edition catbox, latte fileditch, latte catbox, 4.5 fileditch, 4.5

2.5.1.: Filter Spray Edition when i find it in my preset folder. Please help

2.5.: Padoru Edition catbox fileditch

2.3.5: Curry Edition catbox fileditch

2.2: Spooky Edition catbox fileditch

2.1: Avi Edition catbox

2.0: Voodoo Edition catbox fileditch

1.0.3: The one before the big one Edition catbox

1.0.0: I have no idea what I'm doing Edition fileditch

Old Regexes

Credit for delete error goes to Mr. Regex himself. Credit for the remove del regex goes to KnottyDog (i think). End #1 delete del | catbox delete del | fileditch End #2 delete error | catbox delete error | fileditch End #4 & #5 delete repeat | catbox delete repeat | fileditch End #5 & #6 delete curry | catbox delete curry | fileditch

Older Logit Biases

catbox | 0408 fileditch | 0408

catbox| v14 fileditch | v14

My Bots (new)

Once a year, I upload bots. I'll probably upload more of mine. Link to my profile here.