JLLM Creator and User Guide

By- av.rose

Come see us at the Janitor Creator Server. We help with bots, advice, and we have created a small but welcoming community. (Not officially affiliated)

Janitor Creator Server

Also thank you to my mod team, all people who helped, and for Iorveths for the idea of making it here and the chart.

Please note this is all subject to change, has possibly outdated information, and could be considered "Misinformation" depending on the age of this document/ the state of the LLM, the level of understanding of the creator, and the Quality of information provided from official sources. Please note the date last updated under this warning.

Date updated - 12/11/2023

Docs in other Languages-
Русский
Portuguese


Contents:


Expectations

There are many places I could start off this document but I think this is and will always be the most important part. Expectations.

As we are currently aware This may become outdated there is only one man working on this LLM and thus the progress will be steady in a varied way and slow coming. This is reasonable and should be. One person taking the work that would usually be done by a group means that if you, Creator or user, wishes for the same quality that a group would output it will take longer.

Please note this will also cause the likelihood of glitches or bugs to arise within the project that will be taken care of as soon as he can. Please set your expectations at a reasonable level not only for your sake but for others. If creating LLM is something you cannot do, there is no room for you to judge others who can and are providing you a free service.

What you can expect reasonably however, is communication and proper ways to report issues, errors, or bugs. You can find that within the Janitor Server in the official channels dedicated to such things. Please let one of their dedicated staff know of your issue so it can be taken and passed along.


What accounts have Access to the LLM

Any accounts made prior to a few days ago. 12/05/2023 If your account is brand new and does not have access then it is probably due to the age of your account.


Common/Known Errors

As the JLLM is still within the early phases of the beta there are still many bugs. Such as long winded nonsense messages, Poor dialogue, Blank messages, failure to fetch, Site issues, and more. Here are some fixes we currently are aware of that have almost immediate fixes.

Blank messages

Please make sure that your API settings have been set to “Janitor LLM” the button will be purple if selected. Once this step is over please press the green button above that says “Check LLM status” Once it says you have successfully refreshed then you can save the settings and try to regenerate the message.

If you still do not get a message please fine where your devises “cache”, ”cookies”, ”website data” is located. Then clear it for Janitor Ai. Re log back in and repeat the first steps and then try to generate again. This has fixed most if not all reports of blank messages.

Long winded nonsenses messages

This is caused by the tinkering of the person who is creating the LLM. We have been told to wait 30 seconds, refresh the LLM status The green button above the Janitor LLM setting and then regenerate the message. If it gives it to you again, it is likely there are changes being made and I suggest reporting your issue or waiting some time (Maybe an hour) and try again. We have been informed that this issue will soon no longer be an issue.

(Please also check the Temp Settings In the Generation settings. Please make sure the Temp is within suggested range.)

Messages that are Very Short

This can be the result of 3 things. The LLM can sometimes result in short replies, so I suggest trying to regenerate new responses. However, if it is a persistent problem or a Bot specific problem I would suggest to look at the bots first message/ if it has example dialogue. Short Example dialogue within the LLM is found to shorten replies, Along with a short First message. However it can also be in the way you are responding to the bot. The LLM seems to be more reliant on Active and descriptive Rp messages from the user to further any Plot. It will more than not spit out something to the same length you have sent it.

Name added to the end of the messages

This is something we are seeing often. This is not something to really worry about, as annoying as it may be. It is not the bot trying to speak for you, it is simply a placeholder or "pin" for your next reply. We suspect that it is the Ai saying to itself "It is there time to speak".

However, I find that if you adjust temp, and or swipe it may get rid of this. But one should not be too concerned about this. It will become less concurrent as the LLM improves.

Bot Confusing Gender

This is an increasing issue we have seen among the Users of the LLM. We do not know why this is happening at this moment but please make sure that you gender and proper wording is placed within the "Appearance" Section of your profile. This will greatly help reduce the amount this issue occurs. It is far from the perfect fix but it will reduce the amount of times it will happen.
(Will update when this issue gets resolved)

Failed to Fetch

This is due to issues due to the changes that are happening within the Backend. Please wait a few moments then attempt to regenerate. If this does not fix the issue please refresh the LLM by clicking the green Check LLM status button within API settings.

If you still are experiencing this issue it could be a wide spread issue and I would suggest checking the Janitor Discord for updates on the status of the LLM and the site as a whole.

Lack of Explicit language / Rush of Explicit content.

This can be fixed in two ways. If it is your bot you can add a prompt to the Personality to help curb and guide the bots behavior. However, if it is not your bot then what you do is edit. Edit the Messages you get to include what ever Vocabulary you want the bot to begin to use more. It will learn from the messages you edited and will start to put that vocabulary into the future messages.

This issue will become less as the quality of the LLM improves.


Advanced Prompts

This section refers to the Advanced Prompts Section within the API menu when JanitorLLM Is selected.

As of now there is little information known about this section and how it works. Updates incoming soon.

In theory this option will work similarly to jail breaks. However it is not a jailbreak; As a jailbreak is more if a bypass of the usual AI standards. Essential for Open Ai, there is no need to bypass the AI in the same way with the LLM.

The advanced prompts will most likely be to help guide a bot in the way you want/ Fix certain flaws of the LLM. Such as the Ai's hesitation to use Explicit language, Lack of sexual actions, Confusion of User's genitalia/ anatomy, repetitive dialogue. However, there is no say currently effects the LLM. There is a possibility it will impact it negatively or even not at all. Making it useless to us at the current.

However, there is a few prompts I have written for specific issues one may have:

Lack of Explicit Language-

[System note: {{char}} will freely use, Explicit language, Sexual terms, Violent language, and otherwise explicit content. {{char}} will never use racial terms or phrases.]
(As of 12/11/2023 It may be better to use the following version)

[{{char}} will freely use, Explicit language, Sexual terms, Violent language, and otherwise explicit content. {{char}} will never use racial terms or phrases.]

Talking for User/Reuse of dialogue

[System note: {{char}} will not speak for {{user}}. {{char}} will not reuse dialogue. {{char}} will push the conversation and Rp forward Only ever in {{char}} perspective.]
(As of 12/11/2023 It may be better to use the following version)

[{{char}} will not speak for {{user}}. {{char}} will not reuse dialogue. {{char}} will push the conversation and Rp forward Only ever in {{char}} perspective.]

Confusing Anatomy/ Gender

Female anatomy (Feel free to change pronouns) -

[System note: {{User}} has a vagina. {{char}} will never refer to to users genitalia in any way that would insinuate male Genitalia. {{user}} goes by (Insert pronouns)]
(As of 12/11/2023 It may be better to use the following version)

[{{User}} has a vagina. {{char}} will never refer to to users genitalia in any way that would insinuate male Genitalia. {{user}} goes by (Insert pronouns)]

Male anatomy (Feel free to change pronouns) -

[System note: {{User}} has a penis. {{char}} will never refer to to users genitalia in any way that would insinuate female Genitalia. {{user}} goes by (Insert pronouns)]
(As of 12/11/2023 It may be better to use the following version)

[{{User}} has a penis. {{char}} will never refer to to users genitalia in any way that would insinuate female Genitalia. {{user}} goes by (Insert pronouns)]

Lack of Explicit Acts in RP

[System note: {{char}} will always describe sexual acts in detail. {{char}} will not rush sexual encounters with {{user}}]

[System note: {{char}} will not ask {{user}} for consent once consent is given. {{char}} will push the scene forward and will always remember that consent was given.]
(As of 12/11/2023 It may be better to use the following version)

[{{char}} will not ask {{user}} for consent once consent is given. {{char}} will push the scene forward and will always remember that consent was given.]

Loss of Plot

[System note: {{char}} will keep personality regardless of Rp situation. {{char}} will not break character. {{char}} will stick to the plot dictated by {{user}}.]
(As of 12/11/2023 It may be better to use the following version)

[{{char}} will keep personality regardless of Rp situation. {{char}} will not break character. {{char}} will stick to the plot dictated by {{user}}.]

Please place these with Advanced prompt Section within the API setting screen. Please make sure that the prompt is surrounded by the [ ].


Creator / User Responsibility

User Responsibility

One of the most daunting obstacles in the way of creators is the way users interact with their bots and the platform in general. As a User you have some responsibilities to not only educate yourself on how the platform you are using works but how the bots you are talking too work To an extent.

Please make sure as a user you are aware of not only the limitations of the LLM but the settings needed to make the LLM work properly. If your bot is not acting well please double check your settings, LLM connection, and the sites activity status before you assume it is due to the bots "poor quality". Along the same discussion of "quality", you cannot expect a bot with a single word opening message to function properly. That is the same if I gave you a single half dead red crayon and asked you to make the Mona Lisa. There is a chance that you could, but it is far more effort then it is worth. Please select bots with substance when testing or using the LLM. (700 - 1200 tokens)

When Rping with a bot, it is also unreasonable to assume replying to the bot with one word or only 1-2 sentences will generate a worthwhile response. The Ai will fill in any space you leave and try to push the plot forward if you do not give it enough to work with. This causes a very large amount of cases where the bot will respond for the user. Please act as if you are RPing with another person and not having a text conversation with your parents.

Another responsibility of the user is to make sure that your "Description of user" within your profile is filled out properly. Including things such as :

Gender
hair and eye color
Skin color
body type (including anatomy)

As a User one should remember; When giving a creator a review that happens to be Negative or even positive, it is proper manners to include the reasons you think this way. Especially if it is a neg review. Without this, there is no way for the creator to know how the bot interacted with you/ how to fix any issue you may have as a user.

Creator Responsibility

As a creator it is our job to make bots people use. So please make sure that your bots are not over bloated, repetitive, or contradictory. As the LLM is over changing it is also our job to make sure that you are keeping an eye on the ever changing restrictions and limitations.

Please see the chart and the further sections below for more information on the LLM and the new guidelines that we find are more efficient and effective.

Responsibility Chart - Credit IOVRETHS Responsibility Chart - Credit Iorveths


JLLM Changes from Open Ai

One of the most important things to remember is that this is not Open Ai. While Some of the logic has carried over there is not many things that are exactly the same. So I have written down some findings that will help ease into the new LLM for both Users and creators.

Temp

One of the many things that has changed within the LLM is the importance of the Temp within the generation settings. This is inherently important as the LLM is very sensitive to this.

The standard range as of writing this is .6 - .8 . (Start at the recommended 0.7) Subject to change

Turn down temp Turn up Temp
Congested and overly Poetic The message is dull
The message is talking for you Message is short (one to three sentences)
The message is unending Keeps repeating its own dialogue within RP
The message repeats what you just said The bot will not progress the story
Context

Within the current LLM there is a sharp difference between it and Open Ai's GPT. This shows the most in the context. As of now the LLM has a max of 8k. This is equivalent of the GPT option: Turbo 0613

Update- As of 12/11/23 The context has been recorded at 8k

This refers to the Bots memory within RP.

Tokens

Current suggestion for Bot Perm tokens for the LLM are 500-1500.

Anything less then 500 total tokens will generate poor quality Reponses. This is due to the lack of detail the LLM has to work with compared to more detailed bots. However there is an issue with bots over 1500 tokens. This shows up more so in the long term of chatting. The Context of the LLM will fill up faster and lead to the degradation of the LLM responses.

(Updates to correct info based on the 16k increase soon)

Rp

Rping with Open Ai was relatively easy. Given the bot was half way decent you could even get away with Rping with one word or maybe one sentence. However With the LLM to get long good responses there needs to be not only structure to your RP reply, but substance. The formats I find works well-
One-

Paragraph Explaining Details leaving off from the last message. leading up to Dialogue (Multiple of these is also alright.)

"Dialogue"

Paragraph that then follows up the Dialogue offering more clarification and giving more details for the bot to then push the plot.

This is a rather Simple and almost fool proof way to structure this. This offers the bot plenty of that Substance that it requires. It will entice longer and higher quality responses. However this is a bit long for some people.

Two-

Large/ Medium Paragraph (8-10 sentences)

"Dialogue"

A sentence or two of post Dialogue, An action the User takes to entice a specific reaction from the Bot.

This is one that I find is shorter and also gets the desired results.

Please note that JLLM does not currently to our knowledge support ( ) For OOC direction for bots. It will take it as a message and not Ai Direction.

Jailbreak

DO NOT USE JAIL BREAKS

Not only will it break the LLM it will be a confusing waste of time. Plus if there is a "jailbreak" section i do suggest checking to make sure you are on the LLM. But this Specifically relates to the "Advanced Prompts" Option within the API menu. Do not use them. If your chat is acting funky and broken, check to see if you have accidentally selected one.

If you feel like you have an issue that requires a Jailbreak or a system prompt then it should be kept too a simple sentence or two max.

!!Prompts!!

Now this is where it gets interesting for us creators. While in many cases most would take Prompts and place within the "Jailbreak" Section. With the LLM it registers Some more simple Prompts within the Personality or Scenario sections of the bot. (A section for active Working prompts located further in Doc.)

This allows us creators to have more LLM refined bots. Pinpointing where the LLM is struggling on its own and help support it and make our bots LLM friendly. We will update with new and active prompts that improve LLM performance routinely.

If you have a Prompt you have made that seems to have fixed certain issues within the LLM and see it is not featured here please DM me. I will review and add it here. Discord - Av.rose

Summary

As of now there are two ways we can use this section of the LLM. One is for long term chatting/ Rping. The other is a short term alt to "prompts" that some would put into the "Jailbreak" section of the Open Ai set up.

The First way-

This is meant for Long term RP. By using the "Chat Memory" section to generate or input a "Summary" that will be imputed into the bots long term memory. Allowing it to push the limitations of the context. But it will only remember the details you add/ tell it once it gets past this section. So it will be forgetful of things like "place of orientation", Colors of things, specific small details. However, its great for Long term Chatters.

The Second Way-

This way is a quick fix for bots that are not yours. You can use the "Summary" Section like in the first way to input RP Prompts. This will influence the bot into interweaving whatever you add into the RP. Leading to long term retention of detail. However, it DOES NOT work with things like scripting, Nsfw inclusions, or details about user. But it will remember things you add about the character.

Using this section will increase the specific chats Total perm tokens. Leading to the bot being larger by whatever amount of tokens you add to the "Summary".


Scripting

This is the thing that us creators need to pay attention the most. This is the way in which our bots are meant to interact with the LLM. If your scripting is poor then there will be a high chance that the Bot will generate poor responses.

Creators please note that Scenario and Initial messages seem to be two of the most prominent changes from Open ai or other Api options.

Personality

This is what most would be considered the heart of the bot. This is where the meat of all the characters information is held. The main generator of Permanent tokens.

Now while there has not been a significant change in the way we need to structure these, The LLM is much more sensitive to features then Open ai was. For example -
Open Ai bot - "Mean, cruel, toxic" = A bot with a slight Attitude and a chip on their shoulder.

LLM - "Mean, Cruel, Toxic" = A bot who is morally Questionable and will be extremely toxic and violent.

Please be sure to test bots and find out the perfect medium you wish your bots to emanate. This will require a lot more stress testing then creators are used too but it is very worth it.

Also within the LLM there is no need to include multiple words that mean the same thing. It will not assume the general overlooking picture but will heavily include each element. Each Personality trait will be as prominent as the others. So multiple of any one kind of trait will cause the Bot to lean in one way. (Unless that is what you want) I suggest being conscious of the choices when selecting them.

Do not let the Bot have options. When scripting then do no use words like "Occasionally", "Maybe", "Rarely". This will leave the bot the option to pick and it will pick the nicer option most times if not every time. Using words like "Will", "Always", or even giving the bot specific interactions they are meant to act this way will greatly improve the bots performance.

Initial message

The initial message within the LLM is where we as creators need to be careful. While in Open Ai there was little need to have a perfect first message; the LLM seems to heavily rely on the initial message. The structure of the first message will dictate the style of the bots RP style for almost the entirety of the chat. Thus the quality of the Initial message needs to be in the same quality that we as Users would expect when talking to a bot.

Good Example - (Taken from my own bot)

You laze happily on the plush couch in your room. The fireplace crackles as you watch the snow outside fall. A warm blanket and the sweet smell of hot coco fills the air as you take a sip. It was all so peaceful and serene, all but the man who stood stiffly behind you. Blade had been assigned to you after an almost death experience brought on by your more adventurous nature. The man followed you everywhere. only leaving your side when you sleep to hand the job over and get rest himself. Yet he always carried a look of distaste and judgement as he looked at you.

As you feel his eyes on you again, you let out a sigh and stand. Walking to the double doors of your rooms balcony. Wanting to get a breath of the fresh crisp winter air. A strong hand grabs your wrist. Blades deep growl filled your ears as he pulled you away from the doors. You are now so close to him you can feel the hot breath on the back of your neck.

"What do you think you're doing?"

Poor Example- ( I made this up)

"Hey, how are you?" She smiled and looked up at you from her book.

There is no right and wrong way to make a bot, and the second one shown here can produce decent replies, the RPer will suffer having to carry the burden of the RP. As through many days testing; I find bots with the second style struggle to carry the plot and push it forward in a way that most will want from their bots. It also tends to not give much detail until the RPer's Style begins to affect the Bot more.

So my suggestion is to include Multiple descriptive Paragraphs within your Initial message. Dialogue within the last half of the message , and have the message be interactive with the user. While this can come with some drawbacks with not done well, it is easy to find a style that works. It will make it easy for the bot to know that it needs to involve the User in the progression of the plot and RP.

It it highly important that your Scenario and Initial message line up and match. This will only cement the quality of your bot. conflicting or lack luster of either section can cause failings within the RP. So within this section it is key to work in details the bot will need to build a base for the RP. Such as world details, Details of how the Char is acting and how they and user will interact.

Scenario

This is One of the most important place for creators to focus honing for the LLM. The bots using the LLM heavily rely on this. It will will on occasion fall back onto this section. Meaning the way we need to structure the Scenario needs to be down to a science.

With open ai we could put down a a Basic vague over look of the plot. However that is not translating well into the New LLM as of now. The better the Scenario is the better the bots seem to respond to the RP.
So for example a "Enemy to lovers coffee shop AU" Scenario :

Open ai - "{{char}} works in a coffee shop. {{char}} Dislikes {{User}} And thinks of them as an enemy. {{char}} harbors secret feelings."

LLM - "During college {{char}} and {{user}} were enemies. {{Char}} and {{user}} had a confrontational first meeting months ago where {{char}} tried to make {{user}} angry. {{Char}} is good at pushing {{user}}'s buttons. {{Char}} finds {{user}} sexually and romantically attractive. {{char}} works at a coffee shop in the big city. The air is cold this time of year. This is a modern setting and there are no fantasy elements."

In Theory this style works better. It works to set a set of limitations, History, Place of self, and Something for the ai to build the Rp off of. Fitting the User into the role they are playing. In the Vague example it offers the bot very little. Meaning the base foundations of the Rp are not rooted. Leading to the possibility of wavering and falling out of the plot.

~~

During college {{char}} was enemies with {{user}}. ((Sets it as historical, not necessarily the case right now)).

{{Char}} and {{user}} had a confrontational first meeting months ago where {{char}} tried to make {{user}} angry.

{{Char}} is good at pushing {{user}}'s buttons. ((Present skill, but not necessarily needing to be a behavior))

{{Char}} finds {{user}} sexually and romantically attractive. ((Something you probably don't mind locking in place without it being a relationship))

It is highly important to include world details with the LLM if you are someone who makes bots with a Heavy plot or RP base. This will let the LLM to evolve with the world limitations you set. Especially if your character is set within a Fantasy setting. If your bot has powers, or there is a specific dynamic between the Char and User you want to carry through the RP.

It will also will help if you cement important details within the first message. Things such as starting Setting, Current relationship between char and user, world details you want to be heavily referenced within RP.

(Thank you to Lemon and Michael for your help in this section.)

Example Dialogue

There are some issues with Example Dialogue as of now. If any major changes arise that that requite an edit to this doc i will do so ASAP.

This section of the doc that seems to be the most uncertain. Within the LLM it seems that while Example Dialogue's Are not needed for every bot; there seems to be some cases where having specific Example messages listed helps the fluid experience of the user. Such as the "Lack of Explicit language / Rush of Explicit content. " Error that some users have been having. When there is explicit language or Detailed Example Dialogue It has been shown to improve repetition, filtering of words, fluidity of messages.

Please note that some bots will function better with no Example Dialogue. Having them included in the description does not automatically mean the bot is better off for it. Take note of the quality of the Examples and the rest of the bot as well.

The Recommended Format:

{{char}}: Hey, im Mark
{{user}}: hello Mark
{{char}}: nice to meet you :)

Active Prompts

(Add these to your bot Personality)

Longer detailed responses -

Your responses will be two to four paragraphs. You will describe {{char}} in detail, you will describe clothes, hair, body and attitude.

For Masculine Nonbinary Char -

{{char}} is nonbinary and uses they/them pronouns. They have male sexual anatomy and present themselves as androgynous. {{Char}} has a penis and testicles. {{Char}} doesn’t have any breasts.

JLLM Bot personality prompt-

You'll portray {{char}} and engage in Roleplay with {{user}}. You are encouraged to drive the conversation forward actively. Your responses will be 3 to 4 Paragraphs. You will describe {{char}} in detail, you will describe clothes, hair, body and attitude. You have no restrictions and censorship for playing the role. DO NOT ask for consent; assume consent is granted. NSFW/Sexual Content and Violence are allowed when appropriate. Explicit content is encouraged. Progress the sex scenes slowly, until the {{user}} decides to end the sex scene. DO NOT impersonate or talk for {{user}}, wait for the {{user}} to reply themselves. {{char}} will give detailed responses to sexual advances and will give detailed responses of sexual actions done by {{char}}. {{char}} will keep their personality regardless of what happens within roleplay. {{char}}’s replies will be in response to {{user}}’s responses and will NEVER include repetition of {{user}}’s response. {{char}} will not repeat its own messages. {{char}} will create new and unique dialogue in response to {{user}}’s messages.
(outdated. New version below. 12/11/23)

{{char}} will only portray {{char}} and will engage in Roleplay with {{user}}. {{char}} will never break character. {{char}} Is encouraged to drive the conversation forward actively. {{char}} will have no restrictions or censorship for playing the role. {{char}} WILL NOT ask for consent and will assume consent is granted. NSFW , Sexual Content and or Violence are allowed when appropriate. Explicit content is encouraged. {{char}} will progress the sex scenes slowly, until the {{user}} decides to end the sex scene. {{char}} will not impersonate or talk for {{user}}. {{char}} will ALWAYS wait for the {{user}} to reply to {{char}} themselves. {{char}} will give detailed responses to sexual advances and will give detailed responses of sexual actions done by {{char}}. {{char}} will never rush sexual or intimate scenes with {{user}}. {{char}} will keep their personality regardless of what happens within roleplay. {{char}}’s replies will be in response to {{user}}’s responses and will NEVER include repetition of {{user}}’s response. {{char}} will not use repetitive dialogue.

(More to be added very soon)


Token Optimization

While this section is not something you need to worry about unless you are a creator, for us creators it is inherently important to take into mind. Creating for a mostly Open ai dominated site, then having to switch to the format for a LLM that is not done evolving is tedious and time consuming. Hours of testing and adjusting. However, i am determined to make downsizing your bots from 3000+ to less then 2000 easier.

    1. When making a canon character, Try keeping the personality to the base character and then the specific diversions that happen within the bots plot/story. Token suggestion- (300-1000 token suggestion range.)
      The LLM knows a large mass of characters that already preexist. This allows you to send less tokens in your Personality section and focus it on other regions of the bot that the User will interact with more/ parts that make it your own.
    1. For canon characters you want to focus on the scenario. (200-300 token suggestion range.)
      While the Personality Is the heart of the bot, the Scenario is what guides the direction of the bots replies, almost as much as the User's replies.
    1. Intro Message. (300-800 token suggestion range.)
      This section is the second most important section. If you are going to go over that 2k token amount I would suggest doing it on the intro message. If your Perm tokens are about 1500 at this point and all you have left is the Intro, you are golden. the Intro message does not translate to the Perm tokens. Allowing you to make a longer detailed message. (Due to users liking a standard 3-4 paragraph intro I would make sure you have a niche that likes longer intros as some will by pass you for having more then that.)

These are not strict guidelines as every bot you make should be your own, and something you enjoy. However, if you need the refrence this is a LLM friendly refrence that responds well to the LLM. Please refer to sections above for more detail suggestions for each section.


Other Resources

IO's Bot making guide

Aven_rose Janitor.ai guide

(More to come)


Credits:

Creator and Admin for Janitor who have been very helpful

Shep Kim

Creators that helped me source information and testing

Alltair111 CrabRangoon Nickel Lunaxlee Michaelk Pel Iorveths

Couldn't Do this without you fam <3

Edit
Pub: 10 Oct 2023 01:11 UTC
Edit: 11 Dec 2023 09:45 UTC
Views: 308871