What you are describing may be suited for the AI Story Generator assuming you start the run leaving the instructions on total blank and use the “What should happen next?” box as the way to perform actions declaring them on any of the characters that appear on the run.

An even more austere version would be the Prompt Tester assuming you give it the minimal instruction of “Make me a random adventure” or similar and then to continue it paste the contents of the output box in the input box with the extra instruction being “Continue this adventure when <what you are meant to perform next>”.

I should give you a fair warning given how detailed is the story you are putting as a bounty as I imagine you expect that degree of depth in your runs. With the current state of the model, such projects are nearly impossible and would lead to extreme frustration. The best use of the current model are short 1hr-2hr long projects where you aim for a laugh and not quality or consistency.

If you do have the resources, I’d suggest to check the code of the aforementioned generators, as well as the page of the ai-text-plugin to see how to program them and just run your project locally using the resources listed by TRBoom prior. I’ve heard that SillyTavern is also a good alternative. Then again, this all depends on how much hardware you are willing to use.

This is a known “bug” that was seen for long, even from the times of Llama, and one of the reasons why Llama was axed as it did exactly this but more often that the current model. The current one still does it, at the time of release it was rare, but I guess it has become often after several updates. I still find it rare, but the most characters I had at the same time in the current model was five and I find this a rare occurrence, but that is my own experience.

There is a way to dampen this though, which worked in Llama partially. Keep in mind, this is just a “dampen”, as I suspect that due to the way LLMs work, every model in existence would fall to this.

When writing the descriptions of the characters, use three dashes (—) to separate each character. For example:

---
# Character A
<Your description here>
---
# Character B
<Your description here>
---
# Character C
<Your description here>

This is not a fix, but should help it making more rare. Hope this helps though.

Sorry for the late reply. You are kind of the money here. At the time of the original reply, Roleplay 2 was better than Roleplay 1. After so many updates, I can say that actually Roleplay 1 outperforms it due to how some of the problems this new model had are fixed (e.g. cavemen speak)

The reason why often times instructions are longer is to “force” the model to obey them. If you know where the bias of the model is, you can omit certain instructions or just put them in one word, while others that the model “refuses” require lengthy paragraphs before the model reacts.

Under this scope, it is very possible to get a “Roleplay 3” that works flawlessly yet does the same as Roleplay 1 and Roleplay 2 with just a single paragraph worth of text. The problem with doing this however, is that after an update, the bias of the model would change and this would have undesirable effects.

My guess for those two templates to exist today is as a safeguard from the dev, being “Roleplay 1” robust enough when the model is stable, and “Roleplay 2” totally robust even if the model is slightly cookie at the moment.

To know where the bias of the model is, a good experiment is to run a campaign of AI RPG with no prompt, or Story Generator with absolutely nothing, and see what does the model comes up with no instruction. Then start working from that seeing what needs guidance and what works from the get-go.

You can, but it is not straightforward as you may think.

If you press the edit button, on the left hand side of the code, at Line 500 you may find something like this:

return `>>> FULL TEXT of ${letterLabel}: ${messagesText}\n>>> SUMMARY of ${letterLabel}: ${summary}`;

From there forward you see a handful of instructions in plain English that tell the model to generate a summary on the vein of: “Your task is to generate some text and then a ‘SUMMARY’ of that text, and then do that a few more times…” and so on.

Since this instruction is passed in English, the output will be in English as well. If you want to maintain everything in German, you must translate this instruction to German manually.

Now, you’d be surprised but the summaries may not be the culprit of your run being in English randomly, as this principle applies to the how normal instructions are passed, for example, in Line 7291 of the right hand side of the code, you’ll find this:

if(generalWritingInstructions === "@roleplay1") {

And below several instructions in plain English that tell the model how to direct the story. This and several other instructions are passed always each time you press “Send” or the return key, so if you want to be completely sure that your text is never in English, you may need to translate all these instructions as well.

However, something that in the past worked (but I personally have not tested after so many updates this model had undergone so I can’t assure it still works) is that in the Custom Roleplay Style box you can in English write as a prime instruction “The whole text, story, RP MUST be in German (or your desired language)” and it would work without need of translating all.

Granted, this will not change the language of the summaries as the instruction for this is done separately, but it may not affect the output that matters for you.

Hope that helps.

That is a tricky question since it depends a lot on the type of run you are running, and how long it is. Since now the model is (in my opinion) overloaded with new training date, the ideal is to keep all description as terse and succinct as possible. There are however a couple of exceptions to this rule you can use in your advantage.

Ideally, you only want to place all information that is always relevant. For example, your goal, what is the main enemy, inventory if you have it, your current location, etc. However, as described in this guide, you can use it on your advantage to railroad your run into a path and change the setting (i.e., get a breather scene in a war ridden run). You theoretically can put whole character sheets with detailed personalities and all in the Info Tracker. The danger of doing this is that it may end taking precedence over the log itself and the personality of a single character will permeate in the entire world. This is what I describe in the current guide as the “elephant in the room problem” in the “Descriptions and settings” section.

If your run comes from a known IP like World of Warcraft, sometimes all you need to get a more grounded run is to add the magic line Source: World of Warcraft in the Info Tracker and that will automatically load most of the lore for that run in one go without any more tokens. With mistakes inside what is reasonable, but it saves space.

Now, if you want to know why the Info Tracker works that well, is due to how the instructions are passed. Order matters for LLMs, as “Write a story about a cat in the Caribbean” is not the same as telling it “In the Caribbean there is a cat, write his story”. The last part of the input will always have more precedence, so the instructions you place in the Info Tracker are passed AFTER passing the whole log, while the Lore box (the one above the log itself) is passed BEFORE the log.

Under this logic there is a slight potential issue when overloading the Info Tracker, which is that the model will decide to ignore the log and your actual input (i.e., the last thing you said or did to continue the run) in favor to continue the story into something that fits the instructions existing in the info tracker. So while this is indeed a very powerful tool to lead the run, abusing it may cause this unwanted “bug”.

My advise is to place all information that is considered “flavor” in the Lore box, that is, the overall world, character sheets, etc. While using the info tracker to “track” things that are happening at the point you are in the run, keeping it dynamic. You can use it to avoid bad caricaturization, as just go, for example in the case of Cinder “Cinder is a ruthless leader” or similar to provide a nudge while keeping the main information on the Lore box.

There are a million tricks with this model, some new, some inherited from Llama, so again, what is “too much” may become evident if you run in the problem that I just described becomes prevalent, and this depends a lot on the run itself and how long it is.

Hope that helps! 😆

Sorry for the late reply, and thanks actually. My memory is not that great, is just that I try to keep things in order with my own logs to see what is doable and what not. Mainly because this model has undergone more changes that what people report, so the strategies to get a proper run change quickly too (i.e., my old guide is completely useless as a guide today)

My guess with Cinder in your particular case is: yes, it is caricaturization. I have not explained this in the current post but it was in the prior, and the full explanation is as follows:

Behind every generator, there is a set of instructions ready to pass to the model before it generates text. In the case of AI RPG, they go on the lines of “Your task it so create an interactive text adventure where the player is…” and so on. That’s why if you input absolutely nothing and yet press “Generate”, you’ll still get an “adventure” which is the current “default” by the model and a good way to know where its bias is.

Now, after you made your first input, even with instructions and lore, to then press enter and continue the story, you are passing the whole instruction set with your lore PLUS the story at hand. On the code, there is a section that goes along the lines of “This is what happened so far: <your log here>, continue the story”.

If you realize how this goes, the more you advance the story, the more you are feeding the model AI generated text, which will only grow larger and larger to the point that it dominates the custom made text. Causing something like what is shown in this video but with images instead of text.

This is the reason why I call this “caricaturization”. Llama did the same, so all the stories would eventually follow a single format. The current model has more formats, but they are limited, so there is a chance that your setting at that point as “nice enough” that the model decided the Cinder’s behavior would not match the in lore behavior.

No model is safe from this phenomenon due to how instructions are being passed. This is another thing I warn about the current model as this effect was excruciating past the 1Mb mark of the log size at release, while today you can see it happening in a 50kb size log if you are not careful. Again, there are ways to workaround described in this post, so I hope that helps!

… and I took this personally. 😆

Jokes aside, what we all are seeing is something akin to the demonstration in this video:

https://www.youtube.com/shorts/WP5_XJY_P0Q

A theory of why this happens is because the model can’t find a proper way to extrapolate a large large input in a coherent way so it picks random connections causing it to “speak in tongues”.

In the video, the exploit performed is to make the model think it gave as legit advice something outlandish, so it “short circuits” and just gargles random data.

In the case of perchance… I’d be lying if I say I know exactly why it happens now often while in the past this was rare (mind you all, Llama exhibited this too in very niche cases at 20Mb+ and the current model at release I think at 2Mb+, don’t quote me on that, I’m just going from memory), is because again the model is being “hyper-trained” so it fixates only and exclusively on the new training data and not the original data bank it had from fabric. Again, this is just my theory as the opposite could be true, if this model is a “clean” one but with a default language that is not English, it may be struggling with large inputs. Then again I bet more on the earlier given how this model behaved at release compared to today.

Luckily, the workaround is the same as how to cause this artificially: editing the “tongue speak” out and carry on until the model cannot link random parts of its database. It is extremely annoying, but it is not impossible to deal with unless it gets worse and all inputs cause this. In such case, then the model would be broken beyond repair and I hope we don’t get there.

Maybe you are missing the point of the current post, but I’m not asking for Llama, rather warning that updates post November 23rd made the model significantly worse and we are on a path of getting “Llama-like”, as you pointed out in point 2 of your original reply. This is the benchmark I’m using: https://lemmy.world/post/39228619

In my personal opinion, the model was at it’s best around October 10th (based on an old log I have) and around November 23rd (as the linked posts suggests). Ever since it has suffered degradation, and I fear that as time progresses, what today is possible will not be.

To first address the points of your original reply in the numeration you use:

1. Pre October 10th, the model “intelligence” rivaled commercial ChatGPT. Pre November 23rd, we still had great accuracy on borrowed facts and consistency. Today, most things get diluted on the training data. This can be tested with ease with the default character templates (i.e. What you detected as characters randomly gaining tails).
2. This is a consequence of the caricaturization phenomena described here, and it got worst post November. Some would consider this a feature, others a downgrade. I’m on edge on both to be truthful.
3. I suspect this is a problem of over-training the model to “patch” it instead of doing a clean reset. Not much we can do about this as the end users. Workarounds are described above.
4. This was flawless prior to October 10th. Today this “metagaming” stops working after 10 outputs of the introduction of it unless you railroad it.
5. This is just a consequence of your setting. War settings require you to fight something, so the model will provide. The degree of how much you allow depends on your context, if you let the model escalate too much, this will turn into a mess.

Regarding the points of your second reply:

1. This is correct, but past November 23rd, there is a general tone-down in runs that are meant to be by design “dark”. You may get instances of your Char or the world in AI RPG trying to come for an “outlandish peaceful resolution” that makes no sense in the scenario. This is rare, but this is new in the model and shows a trend.
2. Also correct, then again, I warn about the model losing this capacity as the updates happen. As most of the “mystic” concepts are starting to get diluted into “Whimsyland” in the last updates.
3. After November 23rd, this is not correct. Even if you reference a franchise and a specific power/weapon, there are high chances of the model replacing it for something that is more engraved in its training. That being said, we are talking about LLM interpretations of complex definitions, so I don’t expect any degree of accuracy nor is it worth pursuing this when us as the end users can edit the specifics in a run.
4. “Lol u die” was a default in September. By October this was adjusted nicely to then have a dip again and improving back in November. Today it is reasonable, but I still warn towards the model becoming a “tree huger” after a next update due to how data is being handled.

There is a reason I make those posts not only describing problems, but also providing workarounds. Us, as users, have a degree of responsibility with how to direct the model, and there are things that are very possible.

My biggest gripe with the model is, however, that no matter where the bias is, (nice, neutral, or dark) a run comes to “dementia mode” at a lower threshold. I am willing to bet that today it is not possible to get an AI RPG run past the 40 inputs unless one does heavy rerolling and cleaning of the log. This is my fear, that by trying to cater everyone, the model ends producing absolutely nothing.

The current model shows potential, and that’s why it would be sad to see it ending into something that fits no one. Personally, I still believe that a reset with this same model but new training data is required.

Okay, the it is the “luck of the draw”. Keep in mind that this model has its own bias, so the less “evidence” it has, the more it will try to pull you to a state you many not want.

If for some reasons in your logs this happens at the mark of the fourth post, that means that the context you are giving it is 1/4 likely to link your contents to a story you don’t want. Simply erase that message and reroll until you get something that you like. That will reduce the random chance of derailing as you progress.

Keep in mind that this all depends on how much allow the model to modify your run and how many “tools” you give it. A nice character cannot exist in a “violent world”. And since the bias is elsewhere, if you allow an “evil” character and you try to “convert it”, unless you do some heavy workarounds, the model will resist as it will not make sense in the context.

The opposite is true, after the last updates, if your story is too nice oriented, you won’t be able to turn them “organically” into a violent run unless you explicit add the violence. And even then, there are chances of the model returning you to sunshine and rainbows.

Maybe you are trying to go for a realistic story where there is a balance between the two, the problem is that the model will refuse to do this and just stick with the run at hand, so the best approach for this is to actually have the story in mind and only let the model fill the gaps via all the “What happens next” and reminder boxes.

Hope that helps, if you have a more particular problem, do ask, there are a million of workarounds with the current model, and since we were forced to it, best we can do is adapt.

Alright, assuming you are using AI Chat, prompt his character description as follows:

• Name: Rob
• Race: Cyclops
• Appearance: Tall, very slim, blue skin, single magenta colored eye. Lanky frame, pink hands, large head and medium-length auburn hair that partially covers his face. Wears red shorts, a yellow crop-top T-shirt, and dark orange shoes.
• Personality: Outgoing, sociable, easy-going, and enthusiastic.
• Source: The Amazing World of Gumball

The reason why your character is always a bully and you are locked into violence is this:

...and kick him down a manhole, he shows that he can easily forgive by temporarily catching their DVD from the sewers. However, he also demonstrated a sense of entitlement, resentment, and irritability, initially being argumentative...

This new model is not like Llama where you can leave stuff in the background for “later” or as a “explanation”. Whatever that exist on any entry will be used when relevant, and due to the model bias, you are forcing yourself into your character reacting as you described there.

Now, if you REALLY need the “got kicked in the ass and will be irritable and resentful”, you can have it when you deem appropriate by adding it on the personality when you consider it proper, but remember that under the new model lenses, a nice character cannot exist in a world that is not nice to it, locking you in a bad route.

If you really wish to lock the model further, write a small interaction with your Character before letting it loose in the world. You see how the default character templates include some Narrator or Character lines, nothing stops you from giving you a head start where you force a nice interaction of you and the Character before the model takes on.

Hope this helps, and good luck in your runs.

This sub community may be a good fit, but the active community here is small, so no idea if you’d get as much feedback as you desire.

Consider making a repository of your own in services like GitHub or Codeberg and just linking your work in a post, that may make things easier for you.

Good luck. 😄

The current model has several bias and its not perfect, but what you seem to get is an extreme version of problems that are known and there are many workarounds as “chill” runs with “happy and sunshine” characters are possible.

It would help to know exactly what are you prompting to provide aid, as with the default templates (Ike, Li Jung, Quinn, etc) I can’t get exactly what you describe from the get-go.

Correct, that’s what I implied, since otherwise, past the 1Mb you’ll experience “groundhog day” unable to escape the loop no matter what you do.

Now… let me tell you buddy, you just scratched the tip of the iceberg with the model new obsessions. Just to showcase a couple:

• Knuckles turning white (a classic you quoted).
• The ambient smelling like ozone and petrichor (it always rains btw).
• It always smells or tastes of regret and bad decisions.
• The bot or an NPC will always lean to whisper something conspiratorially.
• Eyes gleam with mischief very often.
• Predatory amusement seems to be a normal mood no matter the context.
• Some dialogue constructions are “cursed” as if you let one slide, it will repeat ad nauseam:
  • “Tell me, <text>”
  • “Though I suppose <text>”
• Don’t even let me get started on the “resonance” or “crystallization” rabbit hole…

You are in the money with one thing, all this is product of the training data, and not even the one that comes pre-packed with DeepSeek (I still hold that this is the current model being used, if I’m wrong, I’ll gladly accept the failure on my prediction), this is product of the dataset being used to re-train the model into working for dev’s end. For example, the “knuckles turning white” phrase appeared rarely with the old Llama model, but it was a one in a hundred occurrence as the model didn’t care for that construction and rather focused on a different set of obsessions.

This is a never ending problem with all LLMs though, as in all languages, some constructions are more often than others, and since in both AI Chat and ACC the model is constrained by the “Make a story/roleplay” context, it produces those pseudo-catchphrases incredibly often. In the past we had to deal with “Let’s not get ahead of ourselves” or “We should tread carefully” appearing always no matter the situations, now “knuckles turning white” or similar are the new catchphrases in town.

In an older post I warned about this, since DeepSeek trying to be more “smart” will take everything to face value, so the “correct” answer for many situations tends to be any of these constructions cited, and performing extreme training will yield us a model as dumb and stubborn as Llama was, but with a new set of obsessions plus the inability to move forward which Llama could despite it being exasperating at times. There is progress with the new model, I won’t deny it, but the threshold from were we entered “groundhog day” has been reduced from 1Mb+ to barely 250-500kb and I suspect it will keep reducing if the training is done on top of the existing one, rendering the model pointless for AI Chat, AI RPG or ACC.

Then again, I could be wrong and a future update will allow the context window to hold further as Llama where 15Mb+ was possible and manageable without much maintenance. Some degree of obsession on any LLM is impossible to avoid, what is important is that the model doesn’t turn it into a word salad that goes nowhere. That I think is one of the biggest challenges the development of ai-text-plugin has.

There is a better explanation for the behavior you are experiencing, and yes, it is one if not the biggest hurdle the new model has yet to overcome: You have hit a log long enough that the model is starting to make a word salad of its past inputs as it “inbreeds”.

What I mean by this is something explained before: For generators such as AI Chat and ACC, the input will be mostly 70% AI made and only 30% handwritten (95%-5% in AI RPG which crashes faster), because the whole log is an input for the next output. Of course, the shortest the log is, the less you’ll feel the effect of the model being insufferable because you still have the long instruction block “holding back” the manic behavior.

I agree, this is something that has to be worked on from the development side, otherwise generators such as AI Chat or Story Generator are rendered short-lived as the point of them is to grow progressively, and as today, instability can happen as soon as 150kB-200kB, being significantly lower that what this model was able to hold in the past. However, a temporary fix on our side of things is to just make a “partition” of your log/story. Meaning:

• Plan and start your run as usual.
• Save constantly, monitoring the size of the log.
• When you hit the 100kB mark, try to get to a point where you can “start over” as a point where you can keep moving without requiring the context prior.
• Make a copy, delete all prior to that desired state, load the save and continue pretending that nothing happened.

That will keep the model “fresh” at the cost of losing “memory”, which can be worked around as you can update the bios or instructions which will have better chances of working now under a clean slate.

It is not the best way to work around this, but it is better than wrestling with all the nonsense that the model will produce past the 250kB threshold.

Hope that helps and… also hole that a future update would make the model more stable rather than more unstable. At least something that was fixed and that the dev deserves more credit for making it work, is that at least now the English has improved significantly compared with the first release. In terms of grammar, content and consistency. I know, past the 250kB it is “allegories” or “crazy man ramblings”, but… it is good English! 😅

So… Garth01 called me here, so first of all, thanks for the vote of confidence, buddy! I don’t know if I’m as experienced as you all think but I try my best! 😅

Anyways, about names and why some like the ones repeat a lot. If you are talking about a generator that does not uses the ai-text-plugin like this one, you’ll see on the edit side of things that the names are fixed, passed literally as an array of names as you mentioned:

However, in the case of generators using the ai-text-plugin like ACC when coming up with new characters, or others that write you a long character sheet from a simple input to then make an image or whatnot, that’s because of the training data.

To put it simply, all models require data to work as intended, and depending on such data, it can generate bias. For example, in a random test using the Prompt Tester, you can see this:

You may recognize some of these names depending on what model you use, since as you can see in the prompt, the only “context” given to produce the names is “is for a story”. Changing the context changes the result, as for example, if the context is South America, the model favors “Carlos” or “Maria”, while if the context is Russia, you’ll see it producing “Boris” and “Petrova” often. Note that this is independent of what is the most common name of the region, as the bias is dependent on the training data, which none of us knows what it contains.

It’s the same effect as how the model decides to handle certain situations, for example, if you let it chose the weather, it will pick rain because it has bias towards it. If you let it pick a random encounter against a wild animal, a boar will be more likely. It is not that the model does not recognizes the name, it is just that it has no priority compared to others. Another example would be that even with proper context, you will be extremely unlikely (or even never) get the model to randomly give you the name “Petronilda”, but it recognizes it, as if you ask it about the name, it will give you excruciating detail about its etymology, origin and all.

Contrast to the older model, the new one has more options and is more “creative” as Garth01 mentions. Something many would remember from the old model is that Elara Castellanos and Charles McAllister were omnipresent on all stories to the point that if you dig on the code of some generators such as AI Chat, you’ll see how those were hard banned in the code itself. Then again “more creative” still holds a lot of bias.

Personally, naming is one of the things I don’t let the model pick, because while the new model has more range, it is still limited for many standards and trying to make it “more creative” is a headache that i simply not worth it. Something I did in the past when the old model tried to place a name that was repeated already, was to just change it to something obtained from the Fantasy Name Generator (not by Perchance, this is a third party free service) which contains a large database for pretty much every context you may need.

Hope that helps!

Partially, in the case of Story Generator, since the instruction passed to the LLM is outright “make four paragraphs, less than 400 words” as seen in the code, the output will be abruptly cut. A similar phenomena happens in AI Chat for example, where the order is “write ten paragraphs” but the code makes it so the displayed output is only the first one and the other nine are discarded. A “fun” consequence of this that happened repeatedly in the past with the Llama model and that still happens sometimes, was an output being literally just:

Bot: Bot:

As sometimes the LLM would put the input after a line skip, and the code would throw away the first paragraph due to how the pipeline works. Again, this is a very rare occurrence so it is not worth worrying about it.

Now… there is a bit more on this, but this is just speculation in my side, so take this with a grain of salt since I’m no expert in neural networks, nor in the particularities of some models.

DeepSeek (I still firmly believe that the new model is DeepSeek, even if some argue it may not be) takes some instructions more literally than others. Llama for example had absolutely no regard for length nor consistency in writing style, so you could have one output that was just a line or two, and then the next was a gargantuan thesis that would pretty much advance your story too far from comfort, to then go back to short replies. DeepSeek in contrast looks at the past inputs and tries to gauge how to control lengths. Ironically, something that DeepSeek does in longs runs is try to “extend” the output slowly, hence why if you audit summaries in ACC, AI Chat or AI RPG, you’ll see first very short ones, while later they begin exploding into longer ones until reaching instability and derail in madness.

Also, believe it or not, the model takes all your input, it is not that it doesn’t reach it, it’s just that it decides to ignore it in favor of the context or where your story is because the primary instruction in Story Generator as well as in AI Chat or similar is “continue the story”.

To me here is the biggest difference of the new model and the old one. Llama had almost “written in stone” what a story was meant to be and how to continue it from were you are standing (again, this is speculation from my side having a back catalog of massive logs done in AI Chat and seeing how things progressed there contrast to how they do now). The way Llama “thought” was the following:

• A story must follow the medicine/hero story formula.
• Check the last state and what was prior.
• If there are no stakes, nor clear goal, invent one via a “random happening”.
• If there is a goal but no clear solution, present the “medicine” (random quest, magical MacGuffin, person to go kill).
• If the solution is being worked on, present a method (often “trials to obtain the MacGuffin”)
• If all is solved, then there are no stakes, so rinse and repeat.

While on paper this should work flawlessly as you can put most stories under that formula, it was something that infuriated many users as doing something more “complex” such as adding unforeseen consequences to a method, betrayals, or stories that would not follow that formula was tricky. It was doable, but it required tricking the LLM into a state and making it do your bidding. And as it would require more maintenance and attention to context than just going “auto”, it was something heavily complained in the past.

The new model however, has absolutely no concept of a “formula” for stories, allowing for absolute free-form, making DeepSeek process on how to deal with this task as follows:

• Check the state were the story stands.
• Parse the story prior until there is a precedent on how to continue it.
• If there is none, extrapolate from the data bank.

This is why two things happen: if you are in a state that is vaguely similar to something before, you’ll experience endless deja vu, and if you are faced with the “unknown”, there is the random chance of the LLM to pull a “dark scenario”. Sadly, according to other users, the story itself seems to have precedence over explicit instructions of “no, do this instead”, hence why running in circles forever is a bigger threat and can happen as early as a 20kB log as today (current record of mine at the fourth input in ACC Chloe).

We can hope that this all is improved in the future, but that’s more or less why things happen in my opinion. At least with the new scheme, and seeing how some succeed where I and others fail, I can only deduce that the best way to make the new model “work” is via interpolation, meaning, give it a “target” in the description as “the story purpose is to X get Y, or Z to happen”, so when parsing through the data bank, the LLM will select a similar case as were you are standing and work on it without derailing, granted, this removes completely the “surprise” element, but it’s a decent workaround. Then again, always check the story as is, since the “running in circles forever” is a bigger threat I believe.

Anyways, sorry for the long posts, and good luck in your runs!

Alright, Story Generator is indeed a very tricky one, because even if the model would work as intended, it offers little control.

For the record, don’t trust that much an LLM reply on “why things are how they are” as, for starters, an LLM doesn’t think logically, it just interprets a reply based on the combination of words it faces, and more importantly, the generator itself controls how things are shown and passed, but the LLM just takes one big input and gives one big output, it is not as dynamic as you think it is.

Now, back to Story Generator, something I can advise you to try getting a better experience is to edit in the code from the Perchance side of things Line 21 which restricts to “only four paragraphs” and make it longer to ten or twenty, and also Line 45 which restrict the output to “about only about 400 words”.

The reason for this, is because if the output is short, and the input is gargantuan, the LLM will have a hard time contextualizing what is going on and trying to make something “coherent” within the restrains, this is only true now since the model is still unstable, and in the future it should not be a problem, but for now it may be wise to experiment with longer outputs so the “derailing” is not abrupt.

And another thing that actually will remain true as long as the new model persists: your story as presented IS an input, so before you set instructions, you have to manually edit what you don’t like, or outright prune out a whole section you find out of place. This is because your instructions and the story itself are passed together, so again, if the story is a sad dark one and you insist in the instruction “no, make it happy!”, it won’t happen because the model will look at the story and decide that the only “logical” step is to double down. So yeah, manual work it is. In hindsight, that gives you lee way to see the story itself as an input, as if you manually add a turning point, the LLM will latch on it and work around it instead of following a path and behavior you don’t want in your characters.

Then again, I still think that Story Generator is a really tricky one to work around, I’d put it along with AI Text Adventure which even with the old model would derail into madness as soon as the second input due to how much the context would make the LLM fall into its obsessions quick. Still, with a bit of patience, all can be done, it’s just that it becomes demanding and tiresome, hence why most of us don’t bother anymore in trying fun long runs.

I can’t promise to mod a generator for you now (I owe someone a generator, and time in my side is not nice) but I hope that with those directions you can make the Story Generator give you what you need! Best of luck!

If you use duck.ai, why not Blackbox then or the free version of DeepSeek? Also there are many LLM resources that are free in Helicard. Now, I should warn you, the privacy issue is going to be a lingering demon always. As sad as it may sound, even this site (Lemmy) is heavily compromised on that area, so if privacy is indeed a concern, the best alternative is to go offline.

Then again, I know that hosting an own LLM can be bankrupting expensive, personally it is something I will never be able to do due to economical constrains, so I get it. So… sadly we pay with data, or with cash, one way or another.

Maybe a better idea would be to acquire an API key from a big service such as Gemini or other you may find in HuggingFace with a group of friends to share between many you may trust to cover expenses. Again, I’m just thinking outloud there since I’m unsure what fits your needs.

I would recommend AI Dungeon if the classic version was still available, but it is not, and perhaps you already know of that one and I really don’t like how restrictive it is either.

I don’t know why the heavy backlash on this post. Everyone can ask for an alternative, and it’s not like we are going to pretend that Perchance can make everyone happy.

For alternatives as is… I recall in a post someone mentioning character.ai and Sekai. Personally, I’m not fond of either, as they are very limiting on what can you do and I guess the privacy factor is sketchy on those.

However, while this is going to sound counterintuitive, there is something that Perchance offers us all that no other service offers, which ironically is the answer to what you are looking for:

• Perchance has a whole open source platform for its generators, meaning that it is possible to audit exactly what each generator does and how it passes the information to the model, making anyone able to replicate the exact prompts and pipeline for any LLM you wish to use, locally, with an API key, or using a third party UI.

Meaning that you can turn something as the default “online test for DeepSeek”, “ChatGPT free trial” or “Blackbox AI” into what any of your favorite Perchance generators did. All you need to do is get the prompt and input manually and you are good!

Granted, it is tedious, and for going that route with no coding knowledge, it may be better to try something like SillyTavern, which is just the frontend with no LLM behind.

Then again, while I am also not happy with the update, I’d encourage you and others to be patient. After all, we are given a free LLM to use with almost unlimited tokens, and I believe that the biggest challenge that the dev faces there is not to make the model “literally/story/RP appealing”, but rather “all encompassing while catering to most needs” because the same model that powers ACC, AI Chat, AI RPG and others, is the same model that in other generators has to work as standard AI model that can provide code, information, summaries from documents, etc. So making it work for the generators we use while not destroying its functionality is indeed a heavy challenge.

There was an update very recently that (at least on my side) made the model worse than in the prior (which ironically, made the model work the best at the time, about four days ago). As the dev said in the pinned post, the model is still being worked on, and we are in for a very bumpy ride until things stabilize, but there is at least work being done.

Now, regarding the personality changes, there is something you may want to keep in mind because this may remain true even after the model is perfected: The context of the input has prevalence over descriptions and the recommendation instructions, so it is very difficult to have a character remain happy and joyful if the context forces the model to opt for a more “logical” approach changing it’s character (“logical” in what the LLM training dictates, which often is “moon logic”, but with trial and error it is possible to deduce the word combinations that causes a switch in the wild).

Here is a lengthy guide on the topic. It covers most of the pitfalls you may find. The only thing I believe is no longer a problem (although I may be wrong), is that the “caveman speak” problem seems to be patched already, but again, it is still in the guide in case you run into it and how to restore it. Hope that helps!