Janitor General FAQ and How To Fix
Guide by Shiba. AKA someone who spends too much time on Janitor.
What are tokens?
Tokens are very similar to character counts, though they are counted a bit differently.
https://platform.openai.com/tokenizer is a great resource if you'd like to check how many tokens something is!
On Janitor, you will see "tokens" brought up in several different contexts, primarily:
- Bot's permanent tokens
- Tokens in generation settings
In generation settings, setting your token count to 0 will allow the bot to write as long of a response as it pleases.
Setting the token count to a specific number may cause your bot's responses to get cut off.
Knowing how many permanent tokens your bot has can be a great resource for trouble shooting. You can find this information on the bot's page.
If you would like to learn more about how to properly utilize your bot's tokens, check out this guide.
What does temperature do?
Temperature will affect how creative/nonsensical the bot's responses are.
- On lower temps, you could see things like repetition and bland responses.
- On higher temps, you run the risk of getting gibberish or nonsensical responses.
The recommended temp on Janitor fluctuates. Check the discord and ask what temp people are using!
What is an Advanced Prompt?
An Advanced Prompt is essentially an additional set of instructions you are giving to the AI.
You can find this under the API Settings.
Any instruction you want the bot to follow, you can put here.
AN IMPORTANT NOTE:
LLM's can get confused by negative wording. As an example - "Do NOT speak for user" is a negative prompt. This can confuse LLMs.
You should strive to focus on telling the AI what you WANT them to do, not what you DON'T want them to do!
There are a few advanced prompt guides you can take a look at for ideas!
Mar's Prompts
Astarya's Prompts
Kolach's Prompts
What is an OOC command?
OOC stands for Out Of Character. It was originally used when people were roleplaying with others, and needed to tell the person something outside of the roleplay.
For example: (OOC: Hey I need to go out to dinner, I'll be back later!)
On Janitor, an OOC command can be used as another way to talk directly to the AI in the middle of your roleplay.
You can format this in multiple different ways. A few common examples:
- (OOC: Use more varied language.)
- (System Note: Use more varied language.)
- (Use more varied language.)
Talking for user
There are a few things that can make this worse:
- The bot controls user in it's greeting message.
- The roleplay is in first person.
- "I give you a hug" is an example of first person.
- You are providing short responses that give the bot nothing to work with.
- The bot has been allowed to control user in any previous message.
- Advanced Prompts with negative wording.
- LLMs can struggle to understand negative prompts - "Do not speak as user".
- Don't think about a pink elephant. Did you think about a pink elephant?
There are a few things you can do to combat it:
- Make your messages longer! Give the bot something to work with.
- Use the environment, pick up an apple off the table and start eating it. Be descriptive and imaginative!
- Switch to third person.
- "She wraps his arms around his waist, pulling him into an embrace." is an example of third person.
- Edit any previous response where the bot controls user.
- This cannot be done with intro messages, you will have to try the other methods.
- Use an Advanced Prompt.
- Something as simple as "Only respond as {char}" is enough.
- Make sure your Advanced Prompt uses positive language that encourages what you want the bot to do, NOT what you don't want it to do!
Cut off messages
There are a few things that can make this worse:
- Your internet connection is poor.
- You have your tokens set to a specific number.
- Issues on the backend of Janitor.
- If this is the case, there is nothing you can do until the devs get to it.
There are a few things you can do to combat this:
- Turn off text streaming.
- Set your tokens to 0.
Overly NSFW
There are a few things that can make this worse:
- Your Advanced Prompt includes NSFW instructions.
- The bot's personality includes NSFW.
There are a few things you can do to combat this:
- Edit your Advanced Prompt to not include NSFW.
- Use an Advanced Prompt that targets building up trust and a relationship before delving into NSFW topics.
- Edit any language out of the response that may point to NSFW contexts.
- Sultry, primal, husky, etc are all words that typically correlate with NSFW contexts. Edit these out.
- Use an OOC command reminding the bot to keep things SFW/friendly/platonic.
Repetition
What you can do:
- Edit any previous responses that include this repetition.
- Use an Advanced Prompt that specifies using unique, fresh dialogue.
- Start a new chat.
- You can give the bot a rundown of the previous roleplay. Sometimes it does well with picking it up and continuing the roleplay without issue.
- Use an OOC command to tell the bot to use more varied language.
Responses are too long
What you can do:
- Edit the bot's response to the length you want.
- This includes any previous messages. You cannot edit the greeting message, so you will have to work around this.
- Use an Advanced Prompt.
- Something as simple as "Responses should be a maximum of 3 paragraphs" has helped me in the past.
- Keep your responses the same length you want to see.
- 5 star responses that are the length you want!
- Use an OOC command to tell the bot to keep the responses short.