Mercury FAQ and Troubleshooting Guide

By Glub with help and love from Cheese



Basic Troubleshooting


These are all more or less contributed by the wonderful and lovely cheese AKA HandMistress

  • Temperature: 0.70, higher means "more creative" with word choices, decrease this if it starts acting crazy.
  • Max New Token: 300
  • Max Context Size: 7.9k, decrease this if you get blank message errors or insanity errors.
  • Penalties: 1.10, you can increase/decrease this to suit your needs.
  • Main/System Prompt / Pre-History Prompt:
    • Write {{char}}'s next reply in a fictional role-play between {{char}} and {{user}}. You are required to be proactive, creative, and drive the plot and conversation forward. Be descript and graphic when describing actions, body parts, appearances, clothing, textures, fluids, wants, items, scenery. You are required to be descriptive about {{char}}’s needs, wants, thoughts, and actions. Always end every message from {{char}} with dialogue or actions.
  • Jailbreak / Post-History Prompt:
    • [Never summarize your thoughts. Focus on {{char}}’s perspective. Keep a third person POV in replies.]
    • You can include specific kinks in here as well, as outlined in cheese's rentry here.
    • This is barebones compared to jailbreaks for things like OpenAI and Claude because mythomax is already an uncensored model, so this area functions more as a set of instructions for it to follow rather than intense gaslighting
  • Top P/K: It's fine to keep these at defaults if you don't know what you're doing.
    • Top P: 0.95
      • Higher Top P means it is more "generous' in allowing what flies, basically narrowing in your generations on a very large or very small net. Think of it like closing in on a bell curve. The bigger your Top P, the more you're allowing the LLM to look on the edges of that curve. A Top P of 1.00 looks at the entire net of possibilities, which means it does nothing. Set it to 1.00 to turn it off.
    • Top K: 50
      • A higher Top K allows for more "unlikely" tokens to be strung together. Setting Top K to 100 will make sure that the most likely token is always picked, which makes variations between swipes very small.

Repetition Issues

"My bot keeps saying the same thing over and over! Help! It's just copy pasting straight from my replies/its replies/its definition!"
First off, did you try increasing your penalties? Did you change your context size to 7.5k? If so and you're still having this issue, try decreasing your context size again.

Why? Well... sometimes you have to lobotomize it further to get it to stop being stupid. Basically, if it can't remember the reply it's regurgitating it has to make a new one. Bot can't remember old stupid, has to generate new stupid!


Mercury won't stop speaking for me

First, read this.

Secondly, since you probably didn't read it, the card you're using probably sucks.

The tl;dr of that article is that the MOMENT you let it get away with speaking for {{user}} it will KEEP DOING IT. This INCLUDES any instances in the initial message and example text. There is no quick and easy way to fix this. You have to go hands on.

Basically, you have to edit all instances of "actions" from {{user}} in the AI message to, instead, "reactions" from {{char}}. e.g.

  • {{user}} lifted a hand to {{char}}'s cheek and touched them (user is actively doing something. bad. not good.)
  • {{char}} felt {{user}} touch their cheek (char is reacting to something user is doing. yay! acceptable!)

Boo! that's too much work!! any alternatives? Yes, actually! Find a better card. That's all.


I keep getting straight up nonsense from my bot / it writes too much like Shakespeare on crack.

Decrease your temperature by 0.10 intervals and generate responses until it gets to a readability that you enjoy.


It keeps generating blank messages.

I found that going above a certain Max Context Size will do this. I keep mine around 7.6k for this exact reason


CUDA Error? What's that?

CUDA Error indicates an error with the GPU itself. Not your fault! There's no surefire way to fix this without getting the devs involved, but when it shows up you can either

  • Keep swiping until it goes away
  • Decrease your max context size
  • Wait a bit then come back

Error: character definition too large

This means your character definition total has too many tokens and Mercury can't handle it. What does that mean? Well... first, let's answer what tokens are.

What are tokens?

Basically,

more token = more money (Doesn't quite matter for Mercury because it uses a flat rate)
more word = more token

Context size = basically how much bot can remember at once, including both permanent and temporary tokens. Overloads when the "character definition is too large," i.e. your bot is eating too many tokens, and Mercury is shitting itself because it's too full.

Permanent tokens = run every time you ask bot to generate message. If your bot seems lobotomized and doesn't remember anything you say, it's probably because one of these is eating too many tokens:

  • Main prompt/System prompt / Pre-History prompt
  • Jailbreak / Post-History prompt
  • User persona/appearance
  • Character card (description + personality + scenario)

Temporary tokens = things eventually "forgotten" by the bot as the chat develops to allow for more short term memory.

  • Chat history
  • Example dialogue
  • Initial message

i.e. if your bot has a small token count, it will remember more.

Mercury is still relatively young as far as LLMs go, but many users have noted so far that "smaller" bots run better. there's no hard guidelines for size, but generally it's a good idea to keep your permanent tokens <1k for Mercury. With that out of the way...

The card I am using has too many tokens. What do I do?

Quite honestly, a lot of bots are not well made/optimized. If you really like the gist of a character someone has made but see a lot of "token bloat," you'll likely have to go in there yourself and pull out the weeds for it to run smoothly by overriding the character definitions.

What constitutes a "poorly optimized" bot?

Good question... there are a lot of things, some that can be seen easily at a glance, but some are more egregious and will require more work to get rid of.

First step to diagnosing the card's problems

  • Q: Does it have bits like [OOC: {{char}} will not speak for user] or [System note: {{char}} will not speak for user]?
    • Those things are usually called jailbreaks, which you already have!
    • On chub.venus it's sometimes called the "post history prompt," which, as described above, takes up permanent tokens. What does this mean? well, essentially, you're just adding another jail break that might even conflict with your own, resulting in a weird set of instructions for the AI, as well as token bloat.
    • How to fix: just go ahead and cut those parts out.

It still has a huge amount of tokens. What can I do to slim it down more?

  • Q: Is it an established character? Is any of the information widely known? If you ask a layman (someone who just barely knows about the established character) what they know about the character, will they know it? If so, you can get rid of that information.

    • e.g. {{char}} is Sonic, and the description reads, Sonic is a blue hedgehog who can run at the speed of sound and is enemies with Dr. Eggman.
    • This information is widely known, so you can actually just shorten that to {{char}} is Sonic from the Sonic Franchise, and the bot will go "ah! that Sonic!"
  • Q: does the information look poorly formatted or even... familiar?
    • If you look at it and think, "wait a minute, isn't just a wiki page? Did someone seriously just copy paste it and say "that's a wrap" ? Chances are... Yes, unfortunately. It's a plague upon junior bot makers to do this, which is obviously a lot of token bloat because...
    • The information is usually widely known
    • A lot of the information is unimportant.
    • Try to look through the "wiki entry" type cards to pick out the details you think are actually important... or just find a better made card.
  • Q: does all the information actually matter in terms of rp?
    • Many botmakers will write things for human eyes, which is, quite honestly, not their fault. its pretty much second nature when writing to write something that is enjoyable to read, but the AI can look at that same information and say, "I don't know what to do with this."
      • For example, writing that {{char}} has black and white checkered pants that are slightly frayed at the ends. That's for a human eyeball to read, because you're like "wow, imagery!" but the AI is like "why would i need to know this?" because honestly, when will it need to know your character has specifically black and white checkered pants that are slightly frayed at the ends? it takes up permanent tokens, which means it just runs every. single. message. Every Single Message, the card is pushing the bot to remember, "this character has black and white checkered pants that are slightly frayed at the ends." If you have flavor text like this (things that won't come up in rp), you can probably just gut it, unless you're certain it'd be immersion breaking to do so.
    • Essentially the description section should include things that you think are necessary for it to run every message, things like {{char}} speaks in a cheerful, uplifting tone, and is always kind to others. That's serves as a little nudge saying, "hey, remember this when you write dialogue, okay?"
    • If you know rule "of Chekov's gun," apply it! If you notice that something doesn't come up in rp, delete it. What's it there for if you're not gonna use it?

How do I use Mercury / Can I use Mercury on a front-end that isn't chub.venus.ai ?

Absolutely!

Text Guide
  1. Open https://venus.chub.ai/subscription
  2. Copy the reverse proxy url, https://mercury.chub.ai/mythomax/v1 , which serves as an OpenAI mimic. This is not OpenAI. It merely serves as a way for Mercury to connect to other front-ends by behaving in the same way that OpenAI proxies do.
  3. Click "New Token" and copy the token it generates. Store it somewhere safe, but no worries if you lose it, you can make a new one.
  4. Go to the front-end that you desire and insert the reverse proxy url into the reverse proxy field.
  5. Insert your token as the proxy password and hit save.
  6. Et voila! You are done.
Visual Guide

1
2

Should I get Mars or Mercury? What's the difference between the two?

If you get Mars, you will get access to Mercury as well. If you're not sure whether Mars is worth it, I'd recommend you to check out my log compendium for Mars logs and Mercury logs to compare between the two.

The biggest difference between Mars and Mercury is the parameter size. Mars is 70B whereas Mercury is 13B. "Abstractly, Mercury is a much lighter weight model than Mars. Not as good at logic, worse at memory tasks, but also a very creative writer. Some people prefer it to Mars for certain kinds of roleplay." - quoted from the lovely Joystick_Hero is is much smarter at this sort of thing than I am.


Varies based on situation, but generally you should message/email one of the devs for this sort of question

Lore

Data

I just paid for Mercury and tried using it on the site. It says I don't have tokens, what does this mean?

Basically it means that the servers haven't processed your payment and applied it to your account yet. How long it takes varies, but give it an hour or so if you're having trouble. If you still don't have access, message one of the devs. Make sure to include your chub username, a description of your problem, and proof of purchase.

I'm not happy with the service. Can I get a refund?

Depends on a few things, but the devs have final say, so message them about it

I cancelled my subscription and still got billed. Help?

Message one of the devs about it with a screenshot of your transaction along with an explanation of what happened so they can help you out.

I paid through crypto and I haven't gotten access yet.

The blockchain takes a little while to resolve. Give it a day or two, and if you still don't have access, message one of the devs along with a screenshot of your transaction receipt/proof of payment.

I am unable to pay through crypto or PayPal. Is there another way to pay?

No.

Ok but why....

Fun fact, most payment companies don't like working with sites like chub.ai because of the content hosted there. Stripe used to work with chub but got scared off because of. Y'know. The stuff. It'll probably take a while for new payment options to be added.

Edit
Pub: 26 Aug 2023 04:03 UTC
Edit: 12 Jan 2024 05:35 UTC
Views: 5467