Yukii's Bazaar
Some things I explained in the Discord, some infos I wanted to share, with more or less details. For information, I only have a Mars sub.
- Lazy tips (aka: tips taking two lines or less)
- Why unlimited max tokens is a bad idea ?
- What model for NSFW ?
- What model for NSFW with particular characters ? (Example: monster girls, demons, etc.)
- Asha vs Mixtral ?
- Yukii, why don't you have your context size at the max of what the model can do ?
- Time-lapse, a good idea ? The doom of patterns.
- The good patterns
Lazy tips (aka: tips taking two lines or less)
- Don't set Top P to 0, it will break your chat. If you want to turn it off, set it to 1.
- The minimum length is for Chub' APIs (meaning, Mercury and Mars' models). It isn't supposed to work with other models.
- Don't put jailbreaks in the character's description.
- Actually, don't put any pre history nor post history in your character card at all, unless it is important for you character.
- Really, avoid putting any jailbreak in your character card at all if you want to make it public. It won't work for every model, and may get people banned from OpenAI/Anthropic in the worst case.
- Be aware of your model's specifications (like context size).
- If you are using a pay as you go model, check the price. If you don't, that's an easy way to go broke.
- Don't copy paste a wiki into a character card. It will only eat your tokens.
- DON'T USE W++ ! Even the guy who created it apologized about it, it's outdated, eat your tokens, and doesn't help. You can do lists, but take out the quotation marks and replace the + by commas.
- Stat trackings doesn't work. Even high end models such as Claude 3 and GPT 4 have a lot of difficulties to follow them. They don't know what numbers mean.
- Asterisks for actions. No. Models are trained on plain text, novel style. Asterisks for actions / no quotation marks won't work great. The bot will be able to keep it for a while but will always go back to classic novel style.
Why unlimited max tokens is a bad idea ?
Having a max tokens too high/unlimited can be bad because it gives too much space to the model. Every LLMs use the chat history has a guideline, including the formatting. Long first/example messages = long generated messages.
If the messages in the chat history are too long, and if the max new tokens is too high, the bot will try to generate large messages, no matter what will be inside. For example:
- A ton of paragraphs, that all say the exact same thing but in different ways. (El famoso "Are you ready?" or the variant "Are you sure?")
- The bot may try to "finish" the RP in one go, and talk for user.
- The bot may talk for user.
That's why it's often better to have a real max context size, instead of unlimited.
(And some models, like Chub's just don't support unlimited)
What model for NSFW ?
Honestly, it depends. Personally, I mostly use Asha, only moving to Mixtral when it starts getting too flowery. From what I noticed:
- If you started your chat in a descriptive SFW way, Asha will be very descriptive, a bit flowery and will drag the smut for more messages than Mixtral.
- If you are using a NSFW card, with explicit scenes from the start, Asha will do great too, although Mixtral is more or less on the same level.
- If you like it a bit more crude/explicit, Mixtral is the way to go.
- Although, you can also try Mistral. I actually prefer it over Mixtral, but it needs a bit of context. Definitely better for NSFW than Mixtral imo though .
What model for NSFW with particular characters ? (Example: monster girls, demons, etc.)
Well, I have a demon card, and I played around with all models. So, yeah, it would mostly be Asha.
- Mixtral and Mistral has a tendency to kinda forget the character is a demon after a while, but if you regularly remind it, it's good.
- With Asha, it is better, but again, if it's not NSFW from the start, or with clearly NSFW and very explicit characters, you will get mostly flowery outputs. But if it's NSFW from the start, well... it's extremely steamy.
- Mistral can have a bit of a broken logic, but it's the most descriptive without going into "dance as old as time" kind of descriptive. Probably the best ratio price to quality, since it costs only $5.
It all boils down to how each models handle things:
- Asha follows the chat history A LOT, and will try to get as close to the way the precedent messages have been written as possible (in term of writing style).
- Mixtral follows instructions like a chief (character cards/pre and post history), but I feel like it's quickly bland. It also follows chat history way more than Asha in term of formatting.
- Mistral is a bit of both, but it kinda lacks logic. And some people reported that the character tends to ask a lot about permission, although I've never got it personally.
Also, you can switch models mid-chat. I always do it when I don't like the outputs. For example, you can start with Asha for the logic, and turn to Mistral or Mixtral for the smut.
Asha vs Mixtral ?
Both Asha and Mixtral are in the Mars tier. From my own chats:
Mixtral can become a bit bland after a while, and ends up very easily in the same pattern. It will never speak for user, ad will need to be pushed a lot to talk for NPC or other characters in general. It can also give very good outputs, and I feel like it understands hints easier. Also, it follows instructions a lot.
Asha can be more "creative", but is a bitch, clearly. It's great for groups chats, because it will want to always talk for everybody, and is very fast. You need to advance in the chat or it will take care of it. It can be very.... colorful ? in a way that the messages will often use metaphors, talk about feelings, "time will tell", "two broken souls drifting in the chaos of the world", this kind of shit.
I often repeat it, but you can switch models mid-chat. After a point, when you get too many repetition, or it starts getting bland, or if you are stuck, you can change models. I often do it personally, and it can really spice up your chat.
Yukii, why don't you have your context size at the max of what the model can do ?
I like to stay around 7k tokens for logic, mostly. Even if the models can in theory handle up to 8k tokens, lowering context size can help with logic/memory in the middle of the chat.
Afaik, even if the model read the whole prompt, what is at the beginning and the end get more weight than the rest. What is in the middle can become a bit jumbled. By lowering the max context, things in the middle can be better treated. Again, it's from what I understood of LLMs.
Time-lapse, a good idea ? The doom of patterns.
LLMs (or models, the thing generating the text) are quick to fall into a pattern, meaning the same format of messages or the same words/sentences will appear in also every messages that gets generated.
Some things can get caught by the model easier than other things. For example, if once in a message, there is a dash —, the model will always try to use dashes everywhere, impacting the readability and the quality of the messages.
Doing a time-lapse x days later
, is one of those easy patterns. If you write it once, the model might always add x days later
, x hours later
at every opportunity. And if it does that, your chat will be rushed, like a scenario ending in one messages for example. A list of some "easy patterns" I noticed (with the models I use, might be different for other models):
- Time-laspe [x days/hours later]
- Dashes [—]
- Three dots [...]
- As [in sentences, "as he does something"]
- Exclamations at the beginning of messages [Ah ! Fuck ! Oh ?]
The good patterns
So ! Like we just saw, there are patterns you want to avoid at all costs, because they will literally ruin your chat. But there are patterns you want to have. Well. Not only want, but inevitable. First of all, why do pattern come to life ? Because models, being overglorified autocompletes, use the chat history as a guideline to know how its messages should look like. Well, no. Technically, it's more guessing with probabilities depending how many times something appeared in your text so far, but I already wrote a rentry about the functioning of a LlM, I won't do it again.
Anyways.
You know, more someone do something, more it becomes a habit ? Models have their own little habits, habits that are created with the chat history, until it fill completely their lives and they can think of nothing else. OCD to the extreme. It sees there are a lot of dialogs ? It will always give you a lot of dialogs. It sees the messages are long ? It will always give long message.
It's the Trash In -> Trash Out concept. Whatever the model will get to eat, is whatever it will spit back.
Of course, you can always add something like "Write x number of paragraphs" or whatever, but unless you chat history has X number of paragraphs in every messages from the start, I don't guarantee it will work. Because LLMs are like teenagers who don't listen to instructions, but do what it sees. There are be patterns and shit everywhere ? You will get patterns and shit everywhere, youhou !
PS: That is also why post-history instructions/a jailbreak are so powerful. It's the latest order given to a model, and like a teenager, it might obey to the latest order. Unless you have a nice model that obey (like Mixtral, good kid), instructions that are older will not be obeyed. Because teenagers don't listen to older instructions. At least I didn't.