Welcome to Chatbot AI General #1, the thread for discussing and improving AI pony chatbots.
▶ MLP Bots
https://mlpchag.neocities.org
Spreadsheet (CAI bots + Old repository):
https://docs.google.com/spreadsheets/d/1J7BeqJVDS51cXF8Pgm2YZaFq-Z6ykSJT
SillyTavern website extension: https://github.com/MLPChag/SillyTavern-Chag-Search
Bunch of Character Cards (Descriptions/No Greetings):
https://docs.google.com/spreadsheets/d/1Y6LNOCqAZAWIX-OBEv55HjlzcEdpUh-XbKCpbpL6v5k
CAI bots converted to Tavern: https://files.catbox.moe/ckurq1.zip
Expression packs: https://rentry.org/ChagExpressions
!!!GALLERY!!!: https://drive.google.com/drive/u/2/folders/1Ao-h5HFGMPllSrzSBKM_BvGSiU9f0c2U
▶ How do I start?
1) Select a Frontend
2) Select an AI model
3) Select Jailbreak
4) Select bots
5) Lovemaking with AI mares!
Starting in this hobby can be confusing and difficult. If it’s your first time and you’re lost,
▶ ASK THE THREAD! ◀
Novice-to-advanced guide: https://rentry.org/onrms
▶ SillyTavern (preferred frontend)
https://github.com/SillyTavern/SillyTavern
On Android: https://rentry.org/STAI-Termux
App that voices pony responses in ST: https://drive.google.com/drive/folders/16Ss26VBmgzcSuTGzhaHqRuyVRceTf-YB
▶ More frontends:
Risu: https://risuai.xyz
Agnai: https://agnai.chat
▶ Locals
https://rentry.org/lunarmodelexperiments
>>>/g/lmg/
Mistral Nemo base model fine-tuned on fimfics: https://huggingface.co/Ada321/Nemo_Pony_2/tree/main
▶ Jailbreaks
MLP JB: https://rentry.org/znon7vxe
More JB and guides: https://rentry.org/jb-listing
Hypebots for Tavern: https://rentry.org/pn3hb
▶ Botmaking
Editors: https://agnai.chat/editor
Guides: https://rentry.org/meta_botmaking_list
Advanced: https://rentry.org/AdvancedCardWritingTricks
▶ Miscellaneous
Chag Arts: https://drive.google.com/drive/u/2/folders/1yfajin_hB5rW_jy6WyFjRaRfDA3EUGiJ
https://rentry.org/ChagMiscellaneous
https://rentry.org/ChagArchive
Previous thread: >>42105276
**News:**
- Grok 3 (beta) released on API https://docs.x.ai/docs/models
- OAI? release new, free, test models on OpenRouter: Quasar Alpha and Optimus Alpha
- Chatgpt-4o-latest API updated to match frontend version; native image generation released https://openai.com/index/introducing-4o-image-generation
- Gemini 2.5 Pro Experimental released https://blog.google/technology/google-deepmind/gemini-model-thinking-updates-march-2025/
- Deepseek V3 updated, new weights released https://huggingface.co/deepseek-ai/DeepSeek-V3-0324
/poner/ - Ponies
Ponies and Pony Accessories
218 replies and 90 files omitted.
Poner
No.6926
>>6801
Nice, that's amarezing. I'll certainly be checking this out. I kept using Littlepip and sometimes Blackjack for these, having a proper sim to add mares at times is like a wish come true. Thanks Anon.
Nice, that's amarezing. I'll certainly be checking this out. I kept using Littlepip and sometimes Blackjack for these, having a proper sim to add mares at times is like a wish come true. Thanks Anon.
Poner
No.6939
>>6807
Also you can use catbox to send pngs if they're stripped like they were back in 4chan, but json is fine too.
Also you can use catbox to send pngs if they're stripped like they were back in 4chan, but json is fine too.
Poner
No.7063
>>6979
It also fucks the image up in the process, or at least it did for one of my pictures, I haven't tried any others. I tried to upload it to the booru, which rejected the image with a "width/height must be greater than 0" message. I had to first run it through jpegtran to be able to upload it.
It also fucks the image up in the process, or at least it did for one of my pictures, I haven't tried any others. I tried to upload it to the booru, which rejected the image with a "width/height must be greater than 0" message. I had to first run it through jpegtran to be able to upload it.
Poner
No.7347
>>7189
Depends if they were trained with ponies. Some of the pony models might, but doubt general would have much in local ones
Depends if they were trained with ponies. Some of the pony models might, but doubt general would have much in local ones
>>7186
Thanks you too, anon. My first cards were from mlpchag, so it's nice to see my contribution there.
>>7189
No, overall they are bad with detailed lore, but surprisingly good with crossovers and battle scenes.
I usually use 8B Llama, 12B Mistral, 12B Magnum. All small models are quantized too much, even if they trained with FoE fanfiction, without lorebook they don't understand major events like permanently cloudy skies. But they perfectly understand concept of Fallout and Equestria crossover (I have never met humans in my games).
Instead of token-heavy descriptions and lorebooks I used one quite effective trick for prompting. TF-IDF analysis of FoE compared to 100 random books. This creates a short list of characteristic n-grams (rare phrases of 2-3 words), which works like keys to desired atmosphere, themes and setting. Small models very sensitive to that n-grams regardless of context.
Fallout universe n-grams (basic set):
"nomadic gangs"
"ghouls in ruins"
"Vault Door"
Fallout New Vegas / Western n-grams (expanded set):
"oases of civilization"
"Wild West"
"Rocky Mountains"
Fallout Equestria n-grams (crossover set):
"magical radiation"
"Equestria and Zebrica"
"The Stable"
Military n-grams (make enemies clever):
"Great War"
"small arms and light weapons"
"radio communication"
This phrases must be rare generally, but typical for thematic texts. And then the magic happens, LLM addresses thematic knowledge much more often and in the desired combinations. TF-IDF method allows quantitative estimates for prompting before in-game tests.
[Read more] Thanks you too, anon. My first cards were from mlpchag, so it's nice to see my contribution there.
>>7189
No, overall they are bad with detailed lore, but surprisingly good with crossovers and battle scenes.
I usually use 8B Llama, 12B Mistral, 12B Magnum. All small models are quantized too much, even if they trained with FoE fanfiction, without lorebook they don't understand major events like permanently cloudy skies. But they perfectly understand concept of Fallout and Equestria crossover (I have never met humans in my games).
Instead of token-heavy descriptions and lorebooks I used one quite effective trick for prompting. TF-IDF analysis of FoE compared to 100 random books. This creates a short list of characteristic n-grams (rare phrases of 2-3 words), which works like keys to desired atmosphere, themes and setting. Small models very sensitive to that n-grams regardless of context.
Fallout universe n-grams (basic set):
"nomadic gangs"
"ghouls in ruins"
"Vault Door"
Fallout New Vegas / Western n-grams (expanded set):
"oases of civilization"
"Wild West"
"Rocky Mountains"
Fallout Equestria n-grams (crossover set):
"magical radiation"
"Equestria and Zebrica"
"The Stable"
Military n-grams (make enemies clever):
"Great War"
"small arms and light weapons"
"radio communication"
This phrases must be rare generally, but typical for thematic texts. And then the magic happens, LLM addresses thematic knowledge much more often and in the desired combinations. TF-IDF method allows quantitative estimates for prompting before in-game tests.
Poner
No.7805
>>7803
Works fine on my end. Word of warning on those, I have to wait a flat 2 minutes every message so we may be sticking around here for a bit
Works fine on my end. Word of warning on those, I have to wait a flat 2 minutes every message so we may be sticking around here for a bit
>>362
From the other bunker thread:
>Been on a card creating binge. Have another. This one was inspired by littlewitchers jewel thief Fluttershy.
https://www.characterhub.org/characters/Kadon/the-fillydelphia-four-87f4ba268a9c
>Introducing the Fillydelphia Four, Equestrias best heisting crew. Includes three greetings:
>(you) get taken as a hostage after a job gone wrong
>plan a job with your newly hired crew
>go on your first job as a potential member as the mares test your mettle
https://mlpchag.neocities.org/view?card=Kadon/The%20Fillydelphia%20Four.png
From the other bunker thread:
>Been on a card creating binge. Have another. This one was inspired by littlewitchers jewel thief Fluttershy.
https://www.characterhub.org/characters/Kadon/the-fillydelphia-four-87f4ba268a9c
>Introducing the Fillydelphia Four, Equestrias best heisting crew. Includes three greetings:
>(you) get taken as a hostage after a job gone wrong
>plan a job with your newly hired crew
>go on your first job as a potential member as the mares test your mettle
https://mlpchag.neocities.org/view?card=Kadon/The%20Fillydelphia%20Four.png
>>7828
4chan is sort of back. Captcha can still be a bit fucky, but it's possible to post.
I'll keep lurking both bunker threads to watch activity here, but I’ve got no doubt most of it will return to the main site.
Still, these are good places, and having somewhere a bit less public can be nice for this hobby.
4chan is sort of back. Captcha can still be a bit fucky, but it's possible to post.
I'll keep lurking both bunker threads to watch activity here, but I’ve got no doubt most of it will return to the main site.
Still, these are good places, and having somewhere a bit less public can be nice for this hobby.
Poner
No.7831
It looks like they are still testing things.
Poner
No.7895
For the logs posted on both bunkers, should I just do one recap post with catbox links or just post all of them on different posts? Second one is a bit spammy, I guess, but I like logs and they are important IMO, so I don't care that much.
https://poll.horse/KApDCXNp
https://poll.horse/KApDCXNp
>>7937
>"""main""" one
Reminder that with that you basically mean /mlp/, which is also back.
Honestly I kinda don't want to go back to there, but I'll keep lurking around
>>6801
Instead of a horse, I took the json and decided to use this image instead, felt more fitting for a ambient card
https://ponerpics.org/2100826
Making the town hall center fit in the middle of the rules of third bottom right or so
>"""main""" one
Reminder that with that you basically mean /mlp/, which is also back.
Honestly I kinda don't want to go back to there, but I'll keep lurking around
>>6801
Instead of a horse, I took the json and decided to use this image instead, felt more fitting for a ambient card
https://ponerpics.org/2100826
Making the town hall center fit in the middle of the rules of third bottom right or so
>>7829
>>7937
I'm stay. Just don't want to use ssh-tunnel for 4chan again. Too much efforts, I even don't have expendable vps new. Furthermore /poner/ works just perfectly without captcha and js.
But I still have few cards like Fallout Equestria to translate. And some information about preset-making for our chag-bros.
>>7960
Why not? I like this image too. I just took random mare with duster for card from my /foe/ folder.
>>7937
I'm stay. Just don't want to use ssh-tunnel for 4chan again. Too much efforts, I even don't have expendable vps new. Furthermore /poner/ works just perfectly without captcha and js.
But I still have few cards like Fallout Equestria to translate. And some information about preset-making for our chag-bros.
>>7960
Why not? I like this image too. I just took random mare with duster for card from my /foe/ folder.
>>363
I noticed that /chag/ and /aicg/ anons still don't know about brand new zigger technology, The NoAss Preset:
https://rentry.org/NoAss-GigaSchizoKostyl-Redux
https://rentry.org/noass_ext
English anons sees it just as separate avatars for multi-character cards, but this is much more complex idea.
Our /aicg/ discovered that usual chat-format disturb the creativity and initiative of simulation. Too many limits. Model forbidden to reply as user and his mares in group chat. Model sees post as separated objects and often interpretating user's words as instructions. Overall model sees narration rather like a chat than like a fiction. But best parts of datasets for LLM mostly consist of books!
NoAss, mean 'no assistant'. Model is not assistant. Instead of usual chat this preset allows model to generate endless story. Instead of chat-like posts we use stop sequences for user and other characters. Screenshot is typical example, narration from simulation, then chain of injections from user and his catgirl. But simulation sees user and his catgirl as it's own characters inside story and can completely ignore user's accidental instructions. NoAss preset trick simulation to see user as character inside fiction.
Some schizo-terminology:
- FullAss -- united unificated narrative. Simulation narrate endless story without any control and interruptions by user.
- NoAss -- united not-unificated narrative. Simulation narrate endless story with injections from user and other models.
- HalfAss -- multi-turn unificated narrative. Simulation narrate story, but time to time stops and suggest decisions from user. Like COYA.
- Ass -- multi-turn not-unificated narrative. Simulation narrate story in chat-format and react to posts from user and other models in group chat.
We have already used almost all possible tricks to make Ass better. But ass is just ass. If we look at the LLM logic and output more comprehensively, this is still a lot of possibilities to improvements, if we remove one of deeply hidden unnecessary limits. We can create smooth group chat without separation of narration. We do not need LLM assistance for multi-turn narration. From recent months 2ch /aicg/ slowly drifted from Ass and HalfAss to NoAss format. Finally, anon created Redux Preset for Silly Tavern users. With avatars and much more functions inside.
Try NoAss, bros, this is completely different experience even for well known cards.
[Read more] I noticed that /chag/ and /aicg/ anons still don't know about brand new zigger technology, The NoAss Preset:
https://rentry.org/NoAss-GigaSchizoKostyl-Redux
https://rentry.org/noass_ext
English anons sees it just as separate avatars for multi-character cards, but this is much more complex idea.
Our /aicg/ discovered that usual chat-format disturb the creativity and initiative of simulation. Too many limits. Model forbidden to reply as user and his mares in group chat. Model sees post as separated objects and often interpretating user's words as instructions. Overall model sees narration rather like a chat than like a fiction. But best parts of datasets for LLM mostly consist of books!
NoAss, mean 'no assistant'. Model is not assistant. Instead of usual chat this preset allows model to generate endless story. Instead of chat-like posts we use stop sequences for user and other characters. Screenshot is typical example, narration from simulation, then chain of injections from user and his catgirl. But simulation sees user and his catgirl as it's own characters inside story and can completely ignore user's accidental instructions. NoAss preset trick simulation to see user as character inside fiction.
Some schizo-terminology:
- FullAss -- united unificated narrative. Simulation narrate endless story without any control and interruptions by user.
- NoAss -- united not-unificated narrative. Simulation narrate endless story with injections from user and other models.
- HalfAss -- multi-turn unificated narrative. Simulation narrate story, but time to time stops and suggest decisions from user. Like COYA.
- Ass -- multi-turn not-unificated narrative. Simulation narrate story in chat-format and react to posts from user and other models in group chat.
We have already used almost all possible tricks to make Ass better. But ass is just ass. If we look at the LLM logic and output more comprehensively, this is still a lot of possibilities to improvements, if we remove one of deeply hidden unnecessary limits. We can create smooth group chat without separation of narration. We do not need LLM assistance for multi-turn narration. From recent months 2ch /aicg/ slowly drifted from Ass and HalfAss to NoAss format. Finally, anon created Redux Preset for Silly Tavern users. With avatars and much more functions inside.
Try NoAss, bros, this is completely different experience even for well known cards.
Poner
No.8019
>>8014
>NoAss
Placebo. I tried to use it in non-pony RP, but it persistently put all sorts of mutant-type bullshit or mysticism where it shouldn't be.
From the Russian presets I took remix poluzhop.
>NoAss
Placebo. I tried to use it in non-pony RP, but it persistently put all sorts of mutant-type bullshit or mysticism where it shouldn't be.
From the Russian presets I took remix poluzhop.
>>8011
Alright poner, we better see some zala logs if you're staying, You heard the mods didn't you? Right of passage,
Alright poner, we better see some zala logs if you're staying, You heard the mods didn't you? Right of passage,
>>8063
>You have a right of passage to do blur blur
Do you even know how many guards Zala has fugged? Do you know her lore?
>You have a right of passage to do blur blur
Do you even know how many guards Zala has fugged? Do you know her lore?
Poner
No.8065
Sorry anon, it's the rules.
Poner
No.8066
>>8064
Skill issue, AU it up and take a Zala who's never been with anyone and a Zala who's only been with stallions then convert her to HMD
Skill issue, AU it up and take a Zala who's never been with anyone and a Zala who's only been with stallions then convert her to HMD
Poner
No.8079
>>8062
>Placebo. I tried to use it in non-pony RP, but it persistently put all sorts of mutant-type bullshit or mysticism where it shouldn't be.
NoAss is very demanding to Greeting and formatting. Each card need to rewrite, but script "Rewrite greeting" from NoassExt-utils.json almost useless. In fact, we haven't found consensus about best style of formatting yet, but undoubtedly correct greeting is much more important than with usual ass-prompting.
Try theatrical way of formatting:
1. Act/chapter/location -- first thing in Greeting.
2. Description of location, characters -- second thing.
3. Short example of user's speech and actions -- third thing.
4. Example of location change / event / timeskip -- between.
5. Short example of interactions with another char -- at the end.
>poluzhop
Lol, this literally means HalfAss. I still prefer HalfAss presets too, because of small models, but NoAss looks much more promising, even if it just first steps.
General opinion:
Best for HalfAss -- Gemini
Best for NoAss -- Sonnet
[Read more] >Placebo. I tried to use it in non-pony RP, but it persistently put all sorts of mutant-type bullshit or mysticism where it shouldn't be.
NoAss is very demanding to Greeting and formatting. Each card need to rewrite, but script "Rewrite greeting" from NoassExt-utils.json almost useless. In fact, we haven't found consensus about best style of formatting yet, but undoubtedly correct greeting is much more important than with usual ass-prompting.
Try theatrical way of formatting:
1. Act/chapter/location -- first thing in Greeting.
2. Description of location, characters -- second thing.
3. Short example of user's speech and actions -- third thing.
4. Example of location change / event / timeskip -- between.
5. Short example of interactions with another char -- at the end.
>poluzhop
Lol, this literally means HalfAss. I still prefer HalfAss presets too, because of small models, but NoAss looks much more promising, even if it just first steps.
General opinion:
Best for HalfAss -- Gemini
Best for NoAss -- Sonnet
>>8063
You could just admit that you want new card with Zala, hehe. Okay, challenge accepted.
You know, small models like growing stats. Quite slowly, or not so slowly growing stats.
PS: Speaking seriously, I'm glad, that /mlpol/ still cozy and alive. I missed the birthday event itself, but two years ago found archives and read for hours with wide smile, even without most of images. Mega-impact. Literally.
>Two boards that 4chan did want to end
>Instead became much dear friends
[Read more] You could just admit that you want new card with Zala, hehe. Okay, challenge accepted.
[NAUGHTINESS] - mental state, represents naughty temper. Initially 0%. Increased by daring acts, especially when Zala is in control. Increases her bravery to impressive level. After 25% she is ready for…
You know, small models like growing stats. Quite slowly, or not so slowly growing stats.
PS: Speaking seriously, I'm glad, that /mlpol/ still cozy and alive. I missed the birthday event itself, but two years ago found archives and read for hours with wide smile, even without most of images. Mega-impact. Literally.
>Two boards that 4chan did want to end
>Instead became much dear friends
>>8080
>You could just admit that you want new card with Zala, hehe. Okay, challenge accepted.
I was just thinking the same thing, but for >>8064
She hasn't gotten updated in 2 years and deserves a little better. /mlp/ never went anywhere with co-labbing cards but maybe /poner/ can do better. Fork her, new intros, a rework, not the slop nstdude left us with. I know that anons on /mlp/ turned up their nose at ai assissted intros or ai genned intros. But LLMs are a little better now, what's the opinion on it? I don't think it would be any worse off than the tiny into she already has.
>You could just admit that you want new card with Zala, hehe. Okay, challenge accepted.
I was just thinking the same thing, but for >>8064
She hasn't gotten updated in 2 years and deserves a little better. /mlp/ never went anywhere with co-labbing cards but maybe /poner/ can do better. Fork her, new intros, a rework, not the slop nstdude left us with. I know that anons on /mlp/ turned up their nose at ai assissted intros or ai genned intros. But LLMs are a little better now, what's the opinion on it? I don't think it would be any worse off than the tiny into she already has.
>>8146
The card is old, yeah. It was still Card V1 back then too, so only one greeting was possible. And ntsdude was only doing NSFW.
>What's the opinion on it?
Anons tend to turn their noses up at fully generated AI greetings that are easily identifiable as such. That means:
- Lots of "isms"
- Sentence structures, rhythm, and patterns that are clearly AI-generated
Because they can often influence future outputs and contaminate the interaction right from the start.
Without effort or knowledge of what AI can produce, it's still easy to end up with something subpar like that, even with today's models.
If you do something:
- That is AI-assisted but with human writing as the draft, or
- AI-assisted but not easily identifiable as AI
I don't believe anyone will complain.
[Read more] The card is old, yeah. It was still Card V1 back then too, so only one greeting was possible. And ntsdude was only doing NSFW.
>What's the opinion on it?
Anons tend to turn their noses up at fully generated AI greetings that are easily identifiable as such. That means:
- Lots of "isms"
- Sentence structures, rhythm, and patterns that are clearly AI-generated
Because they can often influence future outputs and contaminate the interaction right from the start.
Without effort or knowledge of what AI can produce, it's still easy to end up with something subpar like that, even with today's models.
If you do something:
- That is AI-assisted but with human writing as the draft, or
- AI-assisted but not easily identifiable as AI
I don't believe anyone will complain.
>>8186
Depends on the model.
You don't see them as much nowadays, but things like:
>testament of
>mix/mixture of X and Y
>too much focus on eyes/gaze, etc.
>mischievous smile or mischievous something
>ministrations
>maybe just maybe
>Somewhere a X (Deepseek)
>your greeting starts with the sun going up, down, illuminating, etc.
For sentence structure, it's mostly about avoiding all your sentences being like:
>{{char}} X "Dialogue." She Y, her voice/eyes Z "Dialogue."
>{{char}}'s eyes/voice/etc. X "Dialogue." She Y, etc.
They are basic sentence structures, so it's normal to have some of them, but certain AIs can spam them a bit too much.
There is also '—' spam on some models etc.
[Read more] Depends on the model.
You don't see them as much nowadays, but things like:
>testament of
>mix/mixture of X and Y
>too much focus on eyes/gaze, etc.
>mischievous smile or mischievous something
>ministrations
>maybe just maybe
>Somewhere a X (Deepseek)
>your greeting starts with the sun going up, down, illuminating, etc.
For sentence structure, it's mostly about avoiding all your sentences being like:
>{{char}} X "Dialogue." She Y, her voice/eyes Z "Dialogue."
>{{char}}'s eyes/voice/etc. X "Dialogue." She Y, etc.
They are basic sentence structures, so it's normal to have some of them, but certain AIs can spam them a bit too much.
There is also '—' spam on some models etc.
>>8167
I strongly prefer to made greeting by myself, despite bad english. Because of bad english.
Card is finished except tests and final editing. My personal filly for cum and amusement, kinky storytelling and silly-tuned political games. Alas, my Zala still too coy for her expected behavior. I mean, to rape poor guards in her particular innocently seductive style. Although I have not yet reached even 50% Naughtiness, at current level she is still too far away even from Cozy level of yangire-psychotype.
Any ideas? Any additions to her lore? Full card, if someone really interested:
https://ponepaste.org/raw/10923
Work in progress, final version for mlpchag after some tests (maybe 2-4 days). In other words I'm first.
I strongly prefer to made greeting by myself, despite bad english. Because of bad english.
Card is finished except tests and final editing. My personal filly for cum and amusement, kinky storytelling and silly-tuned political games. Alas, my Zala still too coy for her expected behavior. I mean, to rape poor guards in her particular innocently seductive style. Although I have not yet reached even 50% Naughtiness, at current level she is still too far away even from Cozy level of yangire-psychotype.
Any ideas? Any additions to her lore? Full card, if someone really interested:
https://ponepaste.org/raw/10923
Work in progress, final version for mlpchag after some tests (maybe 2-4 days). In other words I'm first.
>>8186
>>8191
Actually, this patterns can be found by TF-IDF method, which I mention before.
Every LLM, just like every human being, has very individual writing style (preferences in vocabulary). This style can be detected just with 10k word sample with one of my tools (less than 200 man-hours). For example, this is similarity detection, groups of texts from finfiction-like site. Even when I remove first 50 rarest n-grams from every list (usually names, locations), this works still perfectly and detect author in database of thousands of texts. I think, modern tools can detect authorship just by 1000 word sample or even less.
If you provide at least 10k words output from your model (better FullAss type of narrative from different cards), I can create list of that patterns. This is simple, just bare TF-IDF to compare LLM output to 100 random books. Besides, because it is simple, I'm sure someone made that type of lists already for every model. Try some search first. 4chan archive, github and scientific publications. Keywords "TF-IDF", "2-grams", "3-grams".
>>8191
>There is also '—' spam on some models etc.
Looks like influence of flibusta library dataset. This is our national way to separate complex dialogues and narration. Much better than "spruce forests" I must say.
[Read more] >>8191
Actually, this patterns can be found by TF-IDF method, which I mention before.
Every LLM, just like every human being, has very individual writing style (preferences in vocabulary). This style can be detected just with 10k word sample with one of my tools (less than 200 man-hours). For example, this is similarity detection, groups of texts from finfiction-like site. Even when I remove first 50 rarest n-grams from every list (usually names, locations), this works still perfectly and detect author in database of thousands of texts. I think, modern tools can detect authorship just by 1000 word sample or even less.
If you provide at least 10k words output from your model (better FullAss type of narrative from different cards), I can create list of that patterns. This is simple, just bare TF-IDF to compare LLM output to 100 random books. Besides, because it is simple, I'm sure someone made that type of lists already for every model. Try some search first. 4chan archive, github and scientific publications. Keywords "TF-IDF", "2-grams", "3-grams".
>>8191
>There is also '—' spam on some models etc.
Looks like influence of flibusta library dataset. This is our national way to separate complex dialogues and narration. Much better than "spruce forests" I must say.
Poner
No.8759
>>8690
>vodka runes
This is book titles. Model used central book to find all similar books in library, as result we see all stories of that author and others, who write for that setting. Second example without first 50 terms (character names, locations, setting-related terminology). Gedzerath and PhoenixDev whrite for one setting, but their writhing style completely different. This is different persons.
How it works:
1. Dataset consists of n-grams (words, 2-words, 3-words phrases). Separated lists for each book.
2. Frequently counter make dictionaries from lists. key:value, where value is number of n-grams in list.
3. Percentage: (number of n-gram / sum n-grams in list) = united format for different books.
4. TF-IDF: logarithm (number of n-grams / number of lists with that n-gram) = rarity of n-gram
5. Vector space. Each dictionary as node, rare n-grams as connections, similar values as connection weights.
6. Simplification. Each node to each other node. Only strongest connections will survive after each step.
7. Database: nodes as dictionaries. Keys -- links to other nodes. Values -- pair of n-grams and weights.
8. Summarize: sum of all survived connections as weight between nodes.
9. Format: graph of nodes, connections and weights between them.
10. Output: graphviz.
From first steps this works almost like preparations for new dataset to train LLM, but on "Vector space" stage we used various methods of simplification to study anatomy of the LLM during training. On "Summarize" stage we can see skeleton of LLM, strongest connections and some remnants of internal structure. I used it to make similarity detector for authors in dataset.
This is my own method, nothing professional. I just writefag with some knowledge about coding and old-fashioned syntactic analysis. I even don't know most of modern terminology, just like to do my own research. But results still impressive.
[Read more] >vodka runes
This is book titles. Model used central book to find all similar books in library, as result we see all stories of that author and others, who write for that setting. Second example without first 50 terms (character names, locations, setting-related terminology). Gedzerath and PhoenixDev whrite for one setting, but their writhing style completely different. This is different persons.
How it works:
1. Dataset consists of n-grams (words, 2-words, 3-words phrases). Separated lists for each book.
2. Frequently counter make dictionaries from lists. key:value, where value is number of n-grams in list.
3. Percentage: (number of n-gram / sum n-grams in list) = united format for different books.
4. TF-IDF: logarithm (number of n-grams / number of lists with that n-gram) = rarity of n-gram
5. Vector space. Each dictionary as node, rare n-grams as connections, similar values as connection weights.
6. Simplification. Each node to each other node. Only strongest connections will survive after each step.
7. Database: nodes as dictionaries. Keys -- links to other nodes. Values -- pair of n-grams and weights.
8. Summarize: sum of all survived connections as weight between nodes.
9. Format: graph of nodes, connections and weights between them.
10. Output: graphviz.
From first steps this works almost like preparations for new dataset to train LLM, but on "Vector space" stage we used various methods of simplification to study anatomy of the LLM during training. On "Summarize" stage we can see skeleton of LLM, strongest connections and some remnants of internal structure. I used it to make similarity detector for authors in dataset.
This is my own method, nothing professional. I just writefag with some knowledge about coding and old-fashioned syntactic analysis. I even don't know most of modern terminology, just like to do my own research. But results still impressive.
Poner
No.8850
>>8014
So, turns out I've been doing NoAss for a while. I've always had the irrational distaste for the assistant because I want to make minor edits all over the place, and the perspective shifts trip me up, and make things just not flow.
The techniques in the rentry are more extensive than what I was doing, but they're consistent.
However, I'm pretty certain it won't unlock any power with existing models.
This is evidenced by the effort they have to go through to undo the assistant finetuning, and the loss of zero-shot prompting. Basically, the perfectly fine literaturary ability of the LLM has already been partially lobotomized to make the unnatural user-assistant writing style the default.
A particularly egregious finetune might use this technique to pull the LLM into an entirely distinct, 2nd-person-perspective vectorspace. Most models are not this bad though, especially models that advertise storytelling capabilities, and won't see much change in capability.
It'd be another matter if you were using base models, or managed to find a model that was not tuned as an instruct or chat. These have always been notoriously bad at instruct and chat, exactly because they're naturally NoAss.
[Read more] So, turns out I've been doing NoAss for a while. I've always had the irrational distaste for the assistant because I want to make minor edits all over the place, and the perspective shifts trip me up, and make things just not flow.
The techniques in the rentry are more extensive than what I was doing, but they're consistent.
However, I'm pretty certain it won't unlock any power with existing models.
This is evidenced by the effort they have to go through to undo the assistant finetuning, and the loss of zero-shot prompting. Basically, the perfectly fine literaturary ability of the LLM has already been partially lobotomized to make the unnatural user-assistant writing style the default.
A particularly egregious finetune might use this technique to pull the LLM into an entirely distinct, 2nd-person-perspective vectorspace. Most models are not this bad though, especially models that advertise storytelling capabilities, and won't see much change in capability.
It'd be another matter if you were using base models, or managed to find a model that was not tuned as an instruct or chat. These have always been notoriously bad at instruct and chat, exactly because they're naturally NoAss.
>>8080
Some people like gamification cards too. Like win/fail states What would be something like that for the little zeeb?
Some people like gamification cards too. Like win/fail states What would be something like that for the little zeeb?
1746138655.jpg (117.4 KB, 1000x1000, ced435d5b3e3a1c53d89d44ea9b1df95e4be5bed5f013f8556337d0098321301.jpg)
Deepchink has 2 new models, anyone try them?
Chimera is supposed to be good, but I haven' t had any luck getting it to generate anything.
Chimera is supposed to be good, but I haven' t had any luck getting it to generate anything.
Poner
No.9206
>>8951
Yeah, interfaces. Actually, I hate this part for cards with mares. But this helped a lot for small models, you can reroll bad post earlier, or edit just first part to control following narration. One or two stats + some simple planner helps to control the narration and prevent some loops. Usually models don't have much knowledge about copulation, worse than that most of porn-fics are written extremely bad, at least for my taste. This is very narrow theme for LLM, narrow themes create loops, but even single stat like naughtiness, patience, arousal, stamina from start of replay can add some ground to maneuver.
>>9193
I'm stuck with agnai right now. Maybe for months, but It just cause to make cards better and try some new tricks for old cards. Like "Heroes of Equestria". Simulation with Heroes of Might and Magic gameplay for ancient Equestria setting, during civil war between Celestia and Luna. Never published for English, but I make translation just after Zala.
Yeah, interfaces. Actually, I hate this part for cards with mares. But this helped a lot for small models, you can reroll bad post earlier, or edit just first part to control following narration. One or two stats + some simple planner helps to control the narration and prevent some loops. Usually models don't have much knowledge about copulation, worse than that most of porn-fics are written extremely bad, at least for my taste. This is very narrow theme for LLM, narrow themes create loops, but even single stat like naughtiness, patience, arousal, stamina from start of replay can add some ground to maneuver.
>>9193
I'm stuck with agnai right now. Maybe for months, but It just cause to make cards better and try some new tricks for old cards. Like "Heroes of Equestria". Simulation with Heroes of Might and Magic gameplay for ancient Equestria setting, during civil war between Celestia and Luna. Never published for English, but I make translation just after Zala.
>>362
Heroes of Equestria:
https://raw.githubusercontent.com/Shadybloom/amber-in-the-dark/master/misc/cards/heroes-of-equestria-card.png
https://ponepaste.org/raw/10944
Heroes of Might and Magic in ancient Equestria setting. Random towns and counties, random leaders with random armies, random threat Initially and random events during civil war. Basically wargame with loose rules, but not so bad with numbers.
Actually, sometimes too serious with numbers. Recently, instead of another band of wolf-fillies, or assault of dragon-turtle-fortress we found out that to our tiny Braverock marched two armies, a boar horde of Kaelith The Filly Reaper and multi-headed tentacle daemons, summoned by Nightmare Moon. And they are thousands. But not, we didn't give up. We prepared a plan.
We spread the rumors about treasures, evacuated from Canterlot to Porko-Porko islands. We concentrate all ships, barges, boats from coastal counties to peninsula of Sunset Shores. We lure both hordes there, hoping that they are just kill each others. With help of Darkwood's unicorns and centaurs, under daemonic image our bat-mares bully boars to absolutely ferocious state. We entreat Daybreaker to send her Cloud Riders to call The Storm during battle. Finally, we gather all our forces and evacuated whole peninsula to nearest Darkwoods.
But then both hordes just join forces. They clashes once, even lost few hundreds, but then they decided that ponies are much more tasty. The first plan has failed. We call The Storm immediately, when combined horde started assault the harbor city of Sunset Shores. Under cover of raging rain we attacked their rearguard and supplies. Our pegasus was absolutely badass and heroic, but we lost half of infantry during that insane battle, all battalion of guardmares too, but their sacrifice was not in vain. We poisoned all boar's provision, and then all crops and wells of peninsula. Without food combined horde did not last long and further fate of boars, well, was predictable.
[Read more] Heroes of Equestria:
https://raw.githubusercontent.com/Shadybloom/amber-in-the-dark/master/misc/cards/heroes-of-equestria-card.png
https://ponepaste.org/raw/10944
Heroes of Might and Magic in ancient Equestria setting. Random towns and counties, random leaders with random armies, random threat Initially and random events during civil war. Basically wargame with loose rules, but not so bad with numbers.
Actually, sometimes too serious with numbers. Recently, instead of another band of wolf-fillies, or assault of dragon-turtle-fortress we found out that to our tiny Braverock marched two armies, a boar horde of Kaelith The Filly Reaper and multi-headed tentacle daemons, summoned by Nightmare Moon. And they are thousands. But not, we didn't give up. We prepared a plan.
We spread the rumors about treasures, evacuated from Canterlot to Porko-Porko islands. We concentrate all ships, barges, boats from coastal counties to peninsula of Sunset Shores. We lure both hordes there, hoping that they are just kill each others. With help of Darkwood's unicorns and centaurs, under daemonic image our bat-mares bully boars to absolutely ferocious state. We entreat Daybreaker to send her Cloud Riders to call The Storm during battle. Finally, we gather all our forces and evacuated whole peninsula to nearest Darkwoods.
But then both hordes just join forces. They clashes once, even lost few hundreds, but then they decided that ponies are much more tasty. The first plan has failed. We call The Storm immediately, when combined horde started assault the harbor city of Sunset Shores. Under cover of raging rain we attacked their rearguard and supplies. Our pegasus was absolutely badass and heroic, but we lost half of infantry during that insane battle, all battalion of guardmares too, but their sacrifice was not in vain. We poisoned all boar's provision, and then all crops and wells of peninsula. Without food combined horde did not last long and further fate of boars, well, was predictable.
>>9353
Thanks a lot, anon. Added.
https://mlpchag.neocities.org/view?card=Anonymous/Heroes%20of%20Equestria.png
Will repost it on the 4chan thread so more anons can see it.
Thanks a lot, anon. Added.
https://mlpchag.neocities.org/view?card=Anonymous/Heroes%20of%20Equestria.png
Will repost it on the 4chan thread so more anons can see it.
>>9355
https://desuarchive.org/mlp/thread/42157420/#q42168773
https://files.catbox.moe/9eg45o.txt
I'm approving this edit, forks is good thing. This is more company-tactic oriented version instead of my basic wargame. If it not burdensome for mlpchag owner, I ask to add this edit as alternate "Greeting" to my card.
"Armor +1" is good idea, a strong link to DnD thematic knowledge. I used terms "enchanted armor" and "balanced glaives" in greeting further only because of limitation of context window for reply and my preferences to gather big armies instead of development of small troop.
Overall HoE created as basic simulation card for this kind of edits. I changed a thing or two for every single game. Card tested mainly for local LLM like 8B Llama and 12B Mistral. Some odd phrases, like "manifestation of disloyalty" is specially formed groups of tokens (n-grams) to make a link to thematic knowledge. Useful for small models, but big ones more complex and less sensitive to this trick.
[Read more] https://desuarchive.org/mlp/thread/42157420/#q42168773
https://files.catbox.moe/9eg45o.txt
I'm approving this edit, forks is good thing. This is more company-tactic oriented version instead of my basic wargame. If it not burdensome for mlpchag owner, I ask to add this edit as alternate "Greeting" to my card.
"Armor +1" is good idea, a strong link to DnD thematic knowledge. I used terms "enchanted armor" and "balanced glaives" in greeting further only because of limitation of context window for reply and my preferences to gather big armies instead of development of small troop.
Overall HoE created as basic simulation card for this kind of edits. I changed a thing or two for every single game. Card tested mainly for local LLM like 8B Llama and 12B Mistral. Some odd phrases, like "manifestation of disloyalty" is specially formed groups of tokens (n-grams) to make a link to thematic knowledge. Useful for small models, but big ones more complex and less sensitive to this trick.
>>9355
>>9401
Updated on github. New Greeting from anon + some edit for old one:
https://raw.githubusercontent.com/Shadybloom/amber-in-the-dark/master/misc/cards/heroes-of-equestria-card.png
>>9401
Updated on github. New Greeting from anon + some edit for old one:
https://raw.githubusercontent.com/Shadybloom/amber-in-the-dark/master/misc/cards/heroes-of-equestria-card.png
Poner
No.9425
>>9411
Updated on the website and posted on /mlp/. Again, thanks anon.
https://mlpchag.neocities.org/view?card=Anonymous/Heroes%20of%20Equestria.png
Updated on the website and posted on /mlp/. Again, thanks anon.
https://mlpchag.neocities.org/view?card=Anonymous/Heroes%20of%20Equestria.png
1746730091.png (297.4 KB, 1754x806, ice-queen-stole-my-snowfillies.png)
>>362
Trixie's Snowpity Tale:
https://raw.githubusercontent.com/Shadybloom/amber-in-the-dark/master/misc/cards/trixie-snowpity-tale.png
Classical text quest, magic system with animated items, snowponies, Siberian mythology. Clear goal and methods of investigation (even without user's decisions Trixie can finish story by herself), but random final boss and encounters on the way. Can be CYOA for Trixie, or group chat with your mares.
Quite dangerous adventure, which does not forgive mistakes. Recently during tests for English version my first Trixie waste all mana in desperate fight against icy serpent elemental, just to infuriate foe even more. She manage to blind it and run away to her wagon, but found hersel encircled, deadly tired, and without any kind of weapon. Before she fall unconscious, her beloved teacup, which tried to reach ferry to gather help, was shattering into a dozen porkelain pieces that vanish in the snow.
Trixie's Snowpity Tale:
https://raw.githubusercontent.com/Shadybloom/amber-in-the-dark/master/misc/cards/trixie-snowpity-tale.png
Classical text quest, magic system with animated items, snowponies, Siberian mythology. Clear goal and methods of investigation (even without user's decisions Trixie can finish story by herself), but random final boss and encounters on the way. Can be CYOA for Trixie, or group chat with your mares.
Quite dangerous adventure, which does not forgive mistakes. Recently during tests for English version my first Trixie waste all mana in desperate fight against icy serpent elemental, just to infuriate foe even more. She manage to blind it and run away to her wagon, but found hersel encircled, deadly tired, and without any kind of weapon. Before she fall unconscious, her beloved teacup, which tried to reach ferry to gather help, was shattering into a dozen porkelain pieces that vanish in the snow.
[Reply] [Last 50 Posts] [Last 100 Posts] [Last 200 Posts]
You are viewing older replies. Click [View All] or [Last n Posts] to vew latest posts.
Clicking update will load all posts posted after last post on this page.
Post pagination: [Prev] [1-50] [51-100] [101-150] [151-200] [201-250] [251-268] [Next] [Live (last 50 replies)]
268 replies | 115 files | 121 UUIDs | Page 1
[Add to Thread Watcher]