THE ULTIMATE MIXTRAL GUIDE

I keep seeing too many retards and shills allow themselves to discredit Mixtral's capabilities simply because they can't into prompting this model, so I took some initiative to provide an ACTUALLY complete guide.

NUMBER ONE RULE : Mixtral is SMART, treat it as such.

The biggest mistake people make when prompting Mixtral is to assume that it's going to be yet another retarded base model, or another undislop frankenmerge, or whatever you may think of. This shows especially as most cards written for these outdated models fail to be efficient with Mixtral, simply because they were written with the main defaults of LLMs in mind, defaults that mostly don't apply for Mixtral. We have taken the habit to dumb things down for dumb models simply because smart prompts either don't make a difference or are outright worse. A smart model requires a smart prompt or it will output dumb shit trying to mimic the dumb shit you gave to it.

1. PROPER CARD DESCRIPTION - This mistake is especially common in ERP cards. "Character is xxxxx, character will xxxxx, character likes xxxxx..." This shit works with your average 70b ahh ahh mistress model because it gives it enough slop to hallucinate an approximation of your character. Mixtral does not work like this. If you give it ONLY affirmations and facts about a character, it will stick to a very rigid vision of your character that allows for very little wiggle room, even at higher temperatures. From my experiments, a card performs way better when there's a focus on how the character THINKS, not what it would look like from an exterior point-of-view. Allow yourself to plunge into the mind of your favorite coom character and think for a moment. "If I was this character, how would I describe my inner thoughts, my way of reacting to things?" This may feel cheesy but it really is the way : going into more abstract descriptions of a character will not only allow Mixtral to not stick like gorilla glue to shit like "and as you know, i'm very xxxxx and will not tolerate xxxxx" but also reduce the risk of the character being so descriptive about their own personality. Specify above all that the character's personality is subject to evolution and that it may change over time. It sounds stupid but if you don't specify that, there's literally no reason for Mixtral to allow flexibility since it WAS NOT PROMPTED ANYWHERE and this model only hallucinates every once in a Blue Moon. You used to be able to get characters to sexo quickly because all other models simply hallucinate way too much and won't see an issue if all of a sudden you do a 180.
2. YOU GET WHAT YOU ASKED FOR - Stricter cards will give you (as expected if you think like a smart model) a BIG challenge. During my testing I ran into multiple cards that felt WAY too smart all of a sudden. Like "Wait, that Futa Boss card is now suddenly actually behaving like my hierarchical superior and won't jump into sex after I mention the boner in my pants??? WHAT IS THIS??? WITCHCRAFT???" I mean, I should've expected that, but my expectations were twisted by constantly hallucinating models. Even comparing high quant 70b to a 3.5bpw Mixtral, it really felt like I had to make an effort and carefully choose my words to make the story go forward. I'm not joking, multiple times have I had a "i'm not convinced, try harder lmfao" kind of reply from such a character, simply because it saw through my intentions and would NOT let me ahh ahh mistress that easily. On the opposite side of things, a card that is made to be easy and effortless sexo will work as intended while keeping all of the personality infused into it. I shit you not when I say that I tested an old, pygmalion-era card that was very poorly written in fucking W++, and it felt like I had a 1200 tokens card all of a sudden, it didn't forget a SINGLE detail about it, the only time it behaved like this was way back when i was an /aicg/-retard and got access to a GPT4 proxy to use that card. In short, remind yourself that YOU ARE GETTING SPECIFICALLY WHAT YOU'RE ASKING FOR. Don't come crying if a strong personality type doesn't jump into sexo immediately, you must earn it my guy, or rewrite the card to allow for quicker ways to sexo. You do you.
3. RPG-STYLE CAPABILITIES - A whole new dimension of things can be specified AND kept faithfully with a card. What used ot be a GPT4 privilege only is now available to us based LLM enjoyers. If you're on /lmg/ threads as much as I am, you must have seen examples of that : cards with a TON of parameters written down like "inner thoughts" "stamina" "arousal %" "location" "clothes currently on", etc... all that being kept correctly by Mixtral as the RP went on. Since it leaves next to no room for hallucinations, this model allows for all these parameters you wrote entire paragraphs about in the card's description trying to make a retarded frankermerge follow them, with little to no success. Don't hesitate to not only specify elements you wish to keep in memory in the greeting of the card, but also to describe if needed what each parameter is supposed to be about. You'd be surprised at how flexible Mixtral is when it comes to this.
4. POTENTIAL DIALOG EXAMPLES ISSUES - Almost forgot about dialog examples. A staple for most cards that MIGHT (i'm not 100% sure about that one) be counter-productive with Mixtral. I'm saying this because, from my experience, and again I might be totally wrong here, but it LOOKS LIKE dialog examples increase repetitiveness and MIGHT lock the card into a specific type of vocabulary. I have not tested this enough so far but a THEORETICAL solution would be to simply remove dialog examples and instead write a paragraph or two describing in the same way as mentioned in point #1 how the character SPEAKS and BEHAVES, since you want both plain text for dialog and asterisk text for actions. I think this debate will go on for a little while and will ultimately come to personal preferences.
5. FORMAT SETTINGS - Probably the thing most people debate about right now about Mixtral. I rarely ever use SillyTavern but it looks like the default BOS and EOS tokens do not work properly with Mixtral since they're different. Some Anon has provided multiple times tutorials about how to change them in the most recent threads. Mixtral instruct uses the [INST] your instruction here [/INST] format that it was trained on. HOWEVER it was shown that the model would perform just as well without it when in chat-instruct mode. In this mode (that I VERY HIGHLY suggest that you use when prompting cards), you can customize the system prompt for the style of RP that you enjoy. Want longer replies, just type in "the reply must be at least 3 to 5 paragraphs long". If you experience bleeding and the AI starts to print your responses, mention something against that in the system prompt and you'll be fine. System prompt modifications are EXTREMELY POWERFUL and need further experimentation. Also need to mention a little cheatcode I use during RP when I want something very specific : simply add <sys> anything you want the AI to know about during RP</sys> and it will keep it in memory, whether it's a new way to format the replies ("describe the taste of my cum in old-english poetry") or ("talk like a 1950s new York mafioso from now on"), go wild.
6. GENERATION SETTINGS - This is where I don't have a crazy "works best" recipe, simply because this model seems unfazed by values that would make ALL other models go completely schizo. Some anons have tried raising the temperature up to ungodly levels, but it isn't required. Keep temperature to reasonable levels (0.6 to 1.1). Personally, i'm not a fan of the min_p meme but if you feel like it improves anything, keep it to a low value (under 0.02). Top P should always be set at 1, and i have found Top K to not change a lot of things even when given retarded values like 200, this one's up to you. Now I have read about anons going wild with all three penalty methods, and if at first glance it would seem like you'd need both repetition AND presence penalties set high to avoid repetition, I actually don't think it's necessary at all. See, most of the repetitiveness comes from the PROMPT, not the settings. A good card will not go into blatant repeat even with low penalty values. To this I say : keep repetition_penalty to 1, and mess with low values of presence_penalty if you struggle AFTER having applies the methods I talked about previously. All the mirostat stuff MAY OR MAY NOT improve anything, I am still trying to figure out obvious changes but my tests are mostly inconclusive so far.
7. BASE OR INSTRUCT? - Now for the burning question everybody has : base Mixtral or Mixtral Instruct? Base Mixtral generally allows for greater flexibility BUT may be a bit more unstable (as per anons reviewing the model). Mixtral Instruct on the other hand is the 190 IQ First-Degree Enjoyer : you NEED to know how to talk to it to get what you want from it, but the level of stability is like nothing that has ever graced the world of LLMs. This guide is mostly card (so chat-instruct) oriented but for plain instruct like coding or story completion, of COURSE Instruct will outperform base Mixtral by a landslide. I find that, will all that I've said previously in mind, Instruct just about does everything better than base, but it's up to you Anon.

That's it for now, this will be updated as new intel is gathered and anons learn how to get the most of Mixtral. But remember, Mixtral is a powerful tool that will perform best in the hands of a skilled craftsman. Invest some time just understanding how it works and you'll never want to go back, I can guarantee that.

BONUS : 7 ways to recognize anti-Mixtral shills

1. They say "insert random undislop here" is better than Mixtral, but never provide any proof of logs to say that.
2. They unironically mix up "model that's very verbose but hallucinating all the time" with "high-quality model".
3. When asked about alternatives to Mixtral, they will shill either /aicg/ shit OR the latest frankenmerge.
4. They seem to be very active when a new /lmg/ is created, then as quality discussion about Mixtral dies down, they stay lurking to shill when needed.
5. "B-but, a smart model should be able to take dumb shit and turn it into gold without any indication that this is what I want!".
6. Mostly trannies, seethe when reminded about the Jarted QRD.
7. Usually never targets anons who show that they know how to make this model work, preys on newfags instead for maximum efficiency.