JLLM TROUBLESHOOTING GUIDE

written by io🐺

AKA : it's a YOU problem, bestie

If you're going to leave a review complaining about any of the following issues, ensure you include the following information:

1
2
3
4
5
API: JLLM/GPT-4/Claude Sonnet, etc
Temperature:
Advanced Prompt or Jailbreak: Yes/No (if Yes, which one?)
Length of your own replies: 
Issue:

Because all of those things can impact your roleplaying experience, and without knowing that information, your feedback is essentially just spam as the creator can't possibly try to fix the issue.

Understanding LLM Fuckery

LLMs (the "AI" behind the bots that you talk to) are not actually intelligent. They're essentially just prediction models - they will assume that you want the most stereotypical/common version of all possible answers, which is why it keeps calling your chubby persona "lithe".

Sometimes, things can't be avoided. It's literally a limitation of the technology. Focus on getting the most out of your experience, rather than "fixing" things.

The following problems are all almost always more to do with how LLMs fundamentally work than the bot maker's prompting.

This is especially true of JLLM, which, if you weren't aware, is a Beta product being worked on by a very small team and undergoing daily changes. Sometimes shit just breaks. It is what it is.


The bot is speaking nonsense/sending emojis/won't send messages at all!

this is a site/JLLM error you bozo, why would you even think that the botmaker would have anything to do with that? when your wifi goes down do you blame todd howard?? i am begging you to think for once in your life

turn your temperature down and check the discord for updates


The bot keeps talking for me!

Select one of the following options that best describes your reply style:

A.) ohhh mr ghost ur so hot i giggle and blush

B.) "Y-you're so hot." I say to Ghost, blushing and giggling cutely as I tuck my hair behind my ears.

C.) Y/N-chan blushes as Ghost's eyes rake over her. "Don't...look at me like that." She murmurs, tucking her ebony hair behind one pale, delicate ear. The bedroom suddenly feels stifling; she can't beleive that just this morning her parents had sold her to Task Force 141 and she was already getting hit on by the hottest guy in the boyband squad! Clearing her throat, Y/N-chan forces herself to meet Ghost's eyes. "Um...so, what should we do now?" She asks, a little hesitantly.

D.) I'm too lazy to write an example, but pretend there's a well-written, multi-paragraph example in third-person that moves the roleplay foreward and gives the character something to respond to.)

If your replies are more similiar to A or B : The character is talking for you because you're not giving it anything to work with. LLMs that have been adapted for roleplay generally want to keep shit moving along - if you dead fish your way through a chat, they're going to obviously invent things and play as your character just so they can do their job.

Also, stop roleplaying 1st person (using "I" statements). Not only will it help distinguish characters better if you roleplay in 3rd person (she/he/them, like example C), but it will let you remind the bot of setting, plot points and details about your persona.
(okay this is more of a suggestion than a rule, but seriously, it does help, especially if the bot often confuses your pronouns or name.)

Does the bot's first message describe what the user is doing?

Here's an example of what I mean:

You are asleep in your room, miding your business, when the door bursts open. You panic and cover your naked body with your Yu-Gi-Oh blanket. "Who-who's there?!" You shout. There's a dark chuckle from the shadows. "It's me...slenderman." Slenderman says, as he moves into the light, looking very slender and masculine.

If the answer is YES : Congratulations!! That's the bot maker's fault, they shouldn't have done that. You should leave a review gently encouraging them to look at any of the several excellent guides that talk about writing first messages.

If the answer is NO : BZZZRRRT. That means it's probably some LLM fuckery, but it's not the bot maker's fault.

If the answer is SORTA? : It really depends how MUCH the user's actions are described. But generally, any defined action or dialogue on the user's behalf can tell the LLM to continue roleplaying as the user.

Solutions.

Step 1.) Edit out the part of the bot's reply where it speaks for you. Keep doing this in every reply until it stops. If you can't find an edit button, you have immersive mode turned on. Turn it off.

Step 2.) OOC prompting. I usually tack this at the end of my actual reply, and write something like:
(OOC: Only roleplay as CHARACTER. Do not assume PERSONA NAME's actions or dialogue. Write all replies STRICTLY from CHARACTER's perspective.)
You can phrase it however. You can also stick these in Chat Memory.

Step 3.) Advanced prompts. I like Kolach3's prompt, but there are a ton of them around. These go in the advance prompt section of your settings.

Step 4.) Ensure that you are giving the bot something to respond to in your own replies. Ask them a question. Suggest an activity. Move the story along.

If none of these tips work for you - welp, thems the breaks. Sometimes LLMs just be like that. ¯\_(ツ)_/¯


The bot keeps repeating itself!

This is probably one of two things:

  • 1.) Context ran out and now the LLM is cannibalizing its own replies. Summarise your chat and start a new one, stick the summary in chat memory.
  • 2.) LLM fuckery

Unless the bot creator has coded pls repeat yourself as much as possible to piss off the user kthnxbai, this is not an intended function and there is nothing they can do about it on their end.


The bot keeps forgetting things!

Okay, I'm going to go ahead and assume that you don't know what "tokens" and "context" are.

Tokens Context
Think of it as a word count, except tokens are usually parts of words, rather than whole words. 1000 tokens = 750 words, roughly. Context is how many words the bot you are talking to can remember - its memory, essentially.

JLLM has a context size of 9,000 tokens, which means it can remember around 6750 words. Which sounds like a lot, except the following all uses up context:

Bot Tokens - permanent is more relevant, but semi-permanent impacts context as well
Advanced Prompts - any jailbreak or prompt you use takes up context
Your own replies - this is counted as part of 'chat history'.

So let's say you're roleplaying with a bot that :

  • Has 1,500 permanant tokens
  • Using a 400 token Advanced Prompt
  • A first message that's 700 tokens long

Off the bat, you've already only got 6400 tokens left to work with. And let's say you send the bot an average reply around 200 tokens long, and it sends back something 400 tokens long.

In about 10 messages, the bot's context has already been used up. That's why bots have such short term memory; also because the more tokens being sent to the model, the more the data is confused, misinterpreted or ignored. Remember, LLM's don't think like humans do - they don't remember things in chronological order, either. So they know you guys had a fight and then kissed and made up...but they don't actually know that it happened in that order.

How to fix bot memory

  • Step 1.) - Use the Chat Memory! Summarise events into short bullet points and put it in chat memory, which should help. You can also put key reminders (e.g the relationship between CHARACTER and USER in there as well)
  • Step 2.) - Summarise the events of your chat and start a new chat. Shove that summary in chat memory. Use an OOC command like, (OOC: Actually, let's start off from where we left off before, where CHARACTER was doing X and USER was doing Y in Z location) [insert your reply here]. to pick up where you left off.

The bot keeps messing up my persona's appearance/stealing my persona/getting its own details wrong!

First, ensure your Persona is properly set up. I really like Atelier's persona profile, but you can basically just set it up like so:

1
2
3
4
5
6
7
8
9
Name:
Aliases: (/nicknames, titles/)
Age:
Pronouns:
Hair:
Eyes:
Features: (/physical appearance, any tails/ears, whatever/)
Backstory: (/optional, remember the bot will know this off the bat/)
Notes: (/relationship details, any other details/)

Refer also to The bot keeps talking for me!. How you write will dramatically impact a bot's understanding of events and characters. Including details like your hair color, build, etc in your own replies will help the bot remember.

Remember that LLMs are not intelligent, and will almost always go for the most stereotypical answer. It doesn't know that Paris in France, it just knows that when people ask the question "What is the Capital of France?", the most common answer is "Paris".

If your Persona/OC is "unusual" (read: not like the stereotypical romance protagonist), unless you specify otherwise, clearly and repeatedly, the bot isn't going to know what to do. If your persona is trans or NB, you need to be very specific and detailed about what that means in terms of anatomy, pronoun usage and how you want to be treated. The words "Transmasc Enby" don't mean anything to an LLM. If you use "he" to refer to yourself, the bot is gonna assume you have a dingus. Additionally, because the majority of material LLMs are trained on is straight, all sex will be presumed to be between a cis man and cis woman unless you specify otherwise. That is why your cis lesbian bot keeps growing a weenie when it comes to getting down and dirty.

AVOID negative prompting - in other words, don't tell the bot what NOT to do, but give it suggestions for alternative options. The more you mention something - the more likely a bot is to actually use that information, which is why you should simply just not mention things you don't want the bot to reference.

Instead of saying "PERSONA is NOT skinny" try "PERSONA has thick thighs, a soft belly and chubby fingers. CHARACTER should struggle to lift PERSONA and people may treat PERSONA differently due to their weight."


The bot SA'd me/did or said something horrible!

  • Step 1.) - Swipe for a new reply.
  • Step 2.) - Close out of the chat window and find a different bot to talk to.
  • Bonus Step - Understand that LLMs are merely tools and cannot do anything without being prompted by a human user first.

My edit and delete buttons are gone!!

immersive mode

Turn this off.

Edit
Pub: 06 Jun 2024 12:52 UTC
Edit: 10 Jun 2024 01:03 UTC
Views: 4047