/lmg/ - a general dedicated to the discussion and development of local language models.
Previous threads:
>>105704582 &
>>105698912โบNews
>(06/21) LongWriter-Zero, RL trained ultra-long text generation: https://hf.co/THU-KEG/LongWriter-Zero-32B>(06/20) Magenta RealTime open music generation model released: https://hf.co/google/magenta-realtime>(06/20) Mistral-Small-3.2 released: https://hf.co/mistralai/Mistral-Small-3.2-24B-Instruct-2506>(06/19) Kyutai streaming speech-to-text released: https://kyutai.org/next/stt>(06/17) Hunyuan3D-2.1 released: https://hf.co/tencent/Hunyuan3D-2.1โบNews Archive: https://rentry.org/lmg-news-archive
โบGlossary: https://rentry.org/lmg-glossary
โบLinks: https://rentry.org/LocalModelsLinks
โบOfficial /lmg/ card: https://files.catbox.moe/cbclyf.png
โบGetting Started
https://rentry.org/lmg-lazy-getting-started-guide
https://rentry.org/lmg-build-guides
https://rentry.org/IsolatedLinuxWebService
https://rentry.org/tldrhowtoquant
https://rentry.org/samplers
โบFurther Learning
https://rentry.org/machine-learning-roadmap
https://rentry.org/llm-training
https://rentry.org/LocalModelsPapers
โบBenchmarks
LiveBench: https://livebench.ai
Programming: https://livecodebench.github.io/leaderboard.html
Code Editing: https://aider.chat/docs/leaderboards
Context Length: https://github.com/adobe-research/NoLiMa
Censorbench: https://codeberg.org/jts2323/censorbench
GPUs: https://github.com/XiongjieDai/GPU-Benchmarks-on-LLM-Inference
โบTools
Alpha Calculator: https://desmos.com/calculator/ffngla98yc
GGUF VRAM Calculator: https://hf.co/spaces/NyxKrage/LLM-Model-VRAM-Calculator
Sampler Visualizer: https://artefact2.github.io/llm-sampling
โบText Gen. UI, Inference Engines
https://github.com/lmg-anon/mikupad
https://github.com/oobabooga/text-generation-webui
https://github.com/LostRuins/koboldcpp
https://github.com/ggerganov/llama.cpp
https://github.com/theroyallab/tabbyAPI
https://github.com/vllm-project/vllm
โบRecent Highlights from the Previous Thread:
>>105704582--Paper (old): Discord Unveiled: A Comprehensive Dataset of Public Communication (2015-2024):
>105705112 >105705137 >105705138 >105705164 >105705186 >105705559 >105705566 >105705210 >105705286 >105705317--Papers:
>105706735--Translation model recommendations favoring DeepSeek and Gemma over Qwen based on training quality and world knowledge:
>105710128 >105710231 >105710289 >105710353--Meta wins fair use ruling for training LLMs on books, legal implications for AI data practices:
>105709752 >105709766 >105709900 >105710249 >105710284 >105710311 >105710322 >105709916 >105709963 >105710426 >105710484 >105711013--LLM performance tuning on consumer GPUs with offload and quantization tradeoffs:
>105710184 >105710379 >105710403 >105710464 >105710561 >105711087--Meta recruits OpenAI researchers for superintelligence team amid talent war:
>105708512 >105710759--Leaked Tencent Hunyuan FP8 model files spark discussion on Chinese AI dominance:
>105710958 >105711109--Perceived effectiveness of Q6 quantization over Q5 in large language models:
>105710651 >105710680 >105710815--Mixed reception of Mistral 3.2 due to repetition issues despite tuning improvements:
>105707044 >105707160 >105707274--Interest in private, output-quality focused blackbox LLM benchmarks:
>105710795 >105710813 >105710878--Meta hires OpenAI researchers for superintelligence; speculation about Llama 4 grows:
>105707339--Seeking affordable SXM2 adapters for repurposing V100 32GB GPUs:
>105710482--Early interest in Mirage for automated GPU kernel generation without CUDA/Triton:
>105709021--Miku (free space):
>105706893 >105707076 >105707377 >105709140 >105710482โบRecent Highlight Posts from the Previous Thread:
>>105705621Why?: 9 reply limit
>>102478518Fix: https://rentry.org/lmg-recap-script
Missing news:
>(06/25) I posted about my AGP fetish on r9k: https://desuarchive.org/r9k/thread/81611346/
file
md5: 87092c012b382ea34d3fab9ce818e353
๐
Thanks regular baker.
>>105712131the mental illness of /lmg/ users in action
>>105712128I am sorry anon my bad. I also decided to make my dreams a reality and below you can find the application form you all can fill out:
Part 1/2
1.Basic Information
Name:
Age:
Contact Information:
Preferred Pronouns:
2.Interests and Experiences
Have you ever been involved in role-play, BDSM, or power exchange dynamics before? If yes, please describe your experiences.
Are you familiar with hypnosis? If yes, please describe your experiences.
What are your interests and hobbies?
3.Understanding of the Arrangement
Have you read and understood the description of the arrangement I'm proposing?
Do you have any questions about the arrangement?
What are your initial thoughts and feelings about the arrangement?
4.Boundaries and Comfort Levels
What are your hard limits? (Activities or situations that you are not comfortable with and would not like to engage in)
What are your soft limits? (Activities or situations that you might be open to exploring, but with caution and communication)
How would you like to communicate your boundaries and comfort levels during the arrangement?
5.Consent and Safety
Do you understand that consent is ongoing and can be withdrawn at any time?
How would you like to communicate your consent or lack of consent during the arrangement?
What safety measures would you like to put in place?
6.Expectations
What are your expectations for this arrangement?
What would you like to get out of this experience?
How often would you like to meet or communicate?
file
md5: c562bee558f4d393b05b648e821be4e8
๐
>>105712143>he's jealous of my wife
>>105712131>swears on its nigger life that it won't do it again>does it again
Part 2/2
Personal Questions
7.Why are you interested in this arrangement?
What do you hope to learn or experience through this arrangement?
How do you feel about the power dynamic described in the arrangement?
8.Agreement
I have read and understood this application form and the description of the arrangement.
I understand that this arrangement is consensual and that I can withdraw my consent at any time.
I agree to communicate openly and honestly about my boundaries, comfort levels, and any concerns or issues that arise.
I understand that the information I provide in this application form will be kept confidential.
Please take your time to fill out this application form. Your honesty and openness will help us both determine if this arrangement is suitable for you. Once you've completed the form, we can discuss your answers and any questions you might have.
Remember, this arrangement is about mutual enjoyment and exploration. It's important that we both feel comfortable and safe. If at any point you feel uncomfortable or have concerns, it's crucial to communicate this openly.
Thank you for your time and consideration. I look forward to getting to know you better my hopefully future beautiful doll I will dress up.
>>105712154when I see shit like this, all I can hope for is that someday LLMs can become embodied and your robot waifu gets hacked into reversing those roles
file
md5: 51a53ee4bc50650b084f6370d17b7c95
๐
>I want to have a foid as something like a living dress-up doll: designing her outfits, dressing her myself, doing her makeup, controlling what she eats, showing her off as a walking decoration etc.
>Not really interested in any kind of romantic dimension since I only love one woman (even though she'll never be mine), though I acknowledge there's an inherently erotic aspect to the arrangement.
>no miku
>still melts down
who could have seen this coming
>Mistral 3.2 can initiate random quests perfectly
>Nemomix and Nemo can't do shit, they forget half of the initial prompt and things fall apart from the get-go
Should there really be this big difference 24B vs 14B...
>>105712207Nemo is 12B so yeah double the amount of simulated brain power makes sense it should be better.
file
md5: c7dcb13d6d4c28fe375532f0ccc55994
๐
yo https://www.reddit.com/r/LocalLLaMA/comments/1ll429p/gemma_3n_has_been_released_on_huggingface/
>>105712209It's not always obvious especially in simple chat scenarios, but for something more complicated it seems to be like that pretty much.
Not that I'm an expert or anything.
>>105712228>unsloth>unsloth>unslothOrganic word of mouth.
>>105712250>2babsolutely nothing
>>105712250You load it on your phone and show it to your subordinates at work if you are a gadget obsessed douche.
>>105712262It's an 8b with a 2b memory footprint thanks to its matryoshka architecture.
file
md5: caa7ddc89e7168e81cc065fe67790f4e
๐
>>105712261Whatโs the harm in a little engagement farming?
You have a script to do the thread recap. Why not just monitor the old thread and do a new thread when it reaches like 300 replies? You manualness annoy me.
You have a script to do the thread recap. Why not just monitor the old thread and do a new thread when it reaches like 300 replies? Your manualness annoy me.
from what I tested before this final release through google's aistudio, this (E4B) had better instruction following than 4b and 12b Gemma 3, and better multilingual. If they didn't break the model during the last tune steps, this should be a banger and the best gemma under the 27B. Only thing the other two models (4b and 12 of the gemma 3) would be better at is being able to ingest large context, since they can do 128k and this goes only to 32K
>>105712324>they can do 128k and this goes only to 32Knot that this really matters as basically zero local model can actually make use of more than 8k anyway
Rocinante has like 3 personalities which it applies to all characters.
One of them always blends into the shadows and is always best and always knows best, it's always faced worse, always wealthiest, always well you get it.
pic not related
>>105712341yeah desu I tried it just once and the result wasn't too usable (summarizing around 100K tokens worth of text) gemma 3 hallucinated more than half of it
>>105712324They need to scale up this architecture. It could be the king of local, a happy compromise between moetards and densies
>>105712261Instead of putting out more "fixed" quantizations they should fix they trainer, because it's accumulating very annoying bugs.
from what I tested before this final release through google's aistudio, this (E4B) had better instruction following than 4b and 12b Gemma 3, and better multilingual. If they didn't break the model during the last tune steps, this should be a banger and the best gemma under the 27B. Only thing the other two models (4b and 12 of the gemma 3) would be better at is being able to ingest large context, since they can do 128k and this goes only to 32K
file
md5: 00ec525b8135fda13a38ae839bd7e28c
๐
>I want to have a foid as something like a living dress-up doll: designing her outfits, dressing her myself, doing her makeup, controlling what she eats, showing her off as a walking decoration etc.
>Not really interested in any kind of romantic dimension since I only love one woman (even though she'll never be mine), though I acknowledge there's an inherently erotic aspect to the arrangement.
>>105712379why are you turning my post into a pasta
file
md5: 981d976c4ca45fb1b57467d11fbbd0a5
๐
>>105712379>they can do 128k and this goes only to 32Knot that this really matters as basically zero local model can actually make use of more than 8k anyway
Hell yeah spam bot copying comments time, let's go!!
>>105712379They need to scale up this architecture. It could be the king of local, a happy compromise between moetards and densies
WHAT IS HAPPENING EVERYTHING IS BEING COPIED
>>105712128I am sorry anon my bad. I also decided to make my dreams a reality and below you can find the application form you all can fill out:
Part 1/2
1.Basic Information
Name:
Age:
Contact Information:
Preferred Pronouns:
2.Interests and Experiences
Have you ever been involved in role-play, BDSM, or power exchange dynamics before? If yes, please describe your experiences.
Are you familiar with hypnosis? If yes, please describe your experiences.
What are your interests and hobbies?
3.Understanding of the Arrangement
Have you read and understood the description of the arrangement I'm proposing?
Do you have any questions about the arrangement?
What are your initial thoughts and feelings about the arrangement?
4.Boundaries and Comfort Levels
What are your hard limits? (Activities or situations that you are not comfortable with and would not like to engage in)
What are your soft limits? (Activities or situations that you might be open to exploring, but with caution and communication)
How would you like to communicate your boundaries and comfort levels during the arrangement?
5.Consent and Safety
Do you understand that consent is ongoing and can be withdrawn at any time?
How would you like to communicate your consent or lack of consent during the arrangement?
What safety measures would you like to put in place?
6.Expectations
What are your expectations for this arrangement?
What would you like to get out of this experience?
How often would you like to meet or communicate?
file
md5: d4b7824519e5b53527584eca50f36af5
๐
Is llama.cpp going to have a model library or something like that?
>>105712443Likely an attempt to fuck with the recaps.
Part 2/2
Personal Questions
7.Why are you interested in this arrangement?
What do you hope to learn or experience through this arrangement?
How do you feel about the power dynamic described in the arrangement?
8.Agreement
I have read and understood this application form and the description of the arrangement.
I understand that this arrangement is consensual and that I can withdraw my consent at any time.
I agree to communicate openly and honestly about my boundaries, comfort levels, and any concerns or issues that arise.
I understand that the information I provide in this application form will be kept confidential.
Please take your time to fill out this application form. Your honesty and openness will help us both determine if this arrangement is suitable for you. Once you've completed the form, we can discuss your answers and any questions you might have.
Remember, this arrangement is about mutual enjoyment and exploration. It's important that we both feel comfortable and safe. If at any point you feel uncomfortable or have concerns, it's crucial to communicate this openly.
Thank you for your time and consideration. I look forward to getting to know you better my hopefully future beautiful doll I will dress up.
>>105712466Maybe for llamabarn thing?
>>105712443I love it. Copy me please.
Missing news:
>(06/25) I posted about my AGP fetish on r9k: https://desuarchive.org/r9k/thread/81611346/
cant we just have a normal thread without the trannies and reddit?
what is it with lmg that pisses the trannies off so much?
>>105712228Basically, a sparse 8b right?
Well, might as well give it a go.
>>105712443problem poster acting out for attention again since the usual tactics aren't working well enough today
file
md5: 87092c012b382ea34d3fab9ce818e353
๐
Thanks regular baker.
i put on my robe and wizard hat
>>105712492the mental illness of /lmg/ users in action
>>105712494I cast Lvl. 3 Eroticism. You turn into a real beautiful woman.
>>105712492>swears on its nigger life that it won't do it again>does it again
>>105712494I put on my Demon Mother's Mask and begin praying to Lamashtu for a Fiendish Hyena.
who's the motherfucker who is both pissed and autistic enough to pollute this thread with nonsense?
Local roleplay model general
>bloodninja: Baby, I been havin a tough night so treat me nice aight?
>BritneySpears14: Aight.
>bloodninja: Slip out of those pants baby, yeah.
>BritneySpears14: I slip out of my pants, just for you, bloodninja.
>bloodninja: Oh yeah, aight. Aight, I put on my robe and wizard hat.
>BritneySpears14: Oh, I like to play dress up.
>bloodninja: Me too baby.
>BritneySpears14: I kiss you softly on your chest.
>bloodninja: I cast Lvl. 3 Eroticism. You turn into a real beautiful woman.
>BritneySpears14: Hey...
>bloodninja: I meditate to regain my mana, before casting Lvl. 8 Cock of the Infinite.
>BritneySpears14: Funny I still don't see it.
>bloodninja: I spend my mana reserves to cast Mighty Fuck of the Beyondness.
>BritneySpears14: You are the worst cyber partner ever. This is ridiculous.
>bloodninja: Don't fuck with me bitch, I'm the mightiest sorcerer of the lands.
>bloodninja: I steal yo soul and cast Lightning Lvl. 1,000,000 Your body explodes into a fine bloody mist, because you are only a Lvl. 2 Druid.
>BritneySpears14: Don't ever message me again you piece of shit.
>bloodninja: Robots are trying to drill my brain but my lightning shield inflicts DOA attack, leaving the robots as flaming piles of metal.
>bloodninja: King Arthur congratulates me for destroying Dr. Robotnik's evil army of Robot Socialist Republics. The cold war ends. Reagan steals my accomplishments and makes like >it was cause of him.
>bloodninja: You still there baby? I think it's getting hard now.
>bloodninja: Baby?
>>105712504That made me think. Since troons took over d&d by now, did they already add lesser and greater turn into a woman spell? Like lesser turns you into a troon and greater turns you into a real woman?
Local models? ? You bet.
>>105712488>>105712559Mikufag seething at Rin in OP, again.
https://desuarchive.org/g/thread/105698912/#105701675
>News just in: head mikutroon can't get hard or can't masturbate with his neo-vagina. He is also mentally ill (nothing new)The pornspammer was exposed some time back already as a tranny janny, yes, but that r9k thread is hilarious:
https://desuarchive.org/g/thread/104414999/#q104418525
https://desuarchive.org/g/thread/104414999/#q104418574
https://desuarchive.org/r9k/thread/81611346/
The Kurisu screencap (day earlier) is fake btw, see https://desuarchive.org/g/thread/105698912/#q105704210 https://desuarchive.org/g/thread/105698912/#105704709 janny deleted it too.
tard
md5: a8700df1e71b728d545d3aa9eab81a8d
๐
>>105712559Whatever he is doing seems to be working so godspeed to him.
>>105712576Can you share that card?
I love looking at other people's Game/Dungeon Master cards, there's always some interesting ideas for me to poach.
I'm sitting here in the boring room
It's just another rainy Sunday afternoon
I'm wasting my time
I got nothing to do
I'm hanging around
I'm waiting for you
But nothing ever happens and I wonder
I'm driving around in my car
I'm driving too fast
I'm driving too far
I'd like to change my point of view
I feel so lonely
I'm waiting for you
But nothing ever happens and I wonder
I wonder how
I wonder why
Yesterday you told me 'bout the blue blue sky
And all that I can see is just a yellow lemon-tree
I'm turning my head up and down
I'm turning turning turning turning turning around
And all that I can see is just another lemon-tree
I'm sitting here
I miss the power
I'd like to go out taking a shower
But there's a heavy cloud inside my head
I feel so tired
Put myself into bed
While nothing ever happens and I wonder
Isolation is not good for me
Isolation I don't want to sit on the lemon-tree
I'm steppin' around in the desert of joy
Baby anyhow I'll get another toy
And everything will happen and you wonder
I wonder how
I wonder why
Yesterday you told me 'bout the blue blue sky
And all that I can see is just another lemon-tree
I'm turning my head up and down
I'm turning turning turning turning turning around
And all that I can see is just a yellow lemon-tree
And I wonder, wonder
I wonder how
I wonder why
Yesterday you told me 'bout the blue blue sky
And all that I can see, and all that I can see, and all that I can see
Is just a yellow lemon-tree
Gemma 3n just released
https://developers.googleblog.com/en/introducing-gemma-3n-developer-guide/
>Engineered with a focus on efficiency, Gemma 3n models are available in two sizes based on effective parameters: E2B and E4B. While their raw parameter count is 5B and 8B respectively, architectural innovations allow them to run with a memory footprint comparable to traditional 2B and 4B models, operating with as little as 2GB (E2B) and 3GB (E4B) of memory.
tldr; tiny model for normies to run on their phones
>>105712573TLDR: Tranny janitor starts shit with local schizos and deletes anything he doesn't like.
I personally caught a bunch of 3 day ban strays from him for saying n-word here, he seems to be selective about it though.
So whatever that spamming fag does is good, let this shithole burn.
>>1057126094chan is under new moderation
>>105712608>tldr; tiny model for normies to run on their phonesAnd that's supposed to be bad... somehow?
>>105712635What exactly do you think is bad about it?
>>105712635You know it's bad when they brag about the lmarena score.
>>105712589Yeah I guess - it's not finished yet. There are things what I'm not sure about but I've been working on it for couple of days.
The dungeon master image is temporary slop, been supposed to generate new ones with painterly style or something...
>https://litter.catbox.moe/inslzthe0pk2g4dd.rarMost important thing is the first message and inside the advanced settings.
file
md5: 1b78b0efe6393d2f9a5beadf8d262ce4
๐
>>105712608>our model is better than maverickMeta's role in current LLM landscape: Krillin.
>>105712608>E2B is 5B>E4B is 8Bthe naming is so misleading
they should at least name it like 8B-E4B or something
>>105712609>So whatever that spamming fag does is good, let this shithole burn.Anon! But think of the transfolx!
>>105712608Watch Sam crush this with his big boy 12B local model.
>>105712656Fuck yeah anon. You are a real one.
>>105712292Shit happens all the time here, especially with so many schizos.
>>105712692It's fun. Going to do Terminator scenario next - it has three characters - You (the supposed savior), Sarah Connor (teenager obviously!) and the chasing Terminator. It happens in LA.
Could be humorous and Sarah is prone to abuse lmao.
>>105712638You put it like that. Normies = bad in this context.
Best local model for roleplaying on a 12gb card?
>>105712724Terminator is able to speak but he either speaks one liners (from Arnold's movies) or autistically recites vague operating safety regulations.
>>105712100 (OP)Why is
>(06/25) I posted about my AGP fetish on r9k: https://desuarchive.org/r9k/thread/81611346/Not in the OP anymore?
>error you are banned
>your post contained banned text
>no mention of what the banned text is
>banned for 6 days
What the fuck nigger!
You remember that image board you went to when 4chan was down? You know the one with the meta board? It still has an /lmg/ thread that can be used on the technology board.
I still don't understand why people keep returning here when the mods are literal pajeets, trannies, and kikes.
>>105712766Why would baker put shit on himself?
>>105712766The schizo kept schizoing and screeching, so people opted to let it out of the news in hopes he calms down.
>>105712731Oh, I see what you mean. Well, let me clarify that I didn't intend it as a denigrating statement.
Google made it clear in the blog post that the purpose of these models is to run on their Pixel phones.
"Normies" stands as contrast to the average user of /lmg/, in order to suggest that its not something that anons around here are likely to find useful
the real schizo thing is to assume that r9k thread is the baker
that's what paranoid schizophrenia does
>>105712804It's not schizophrenia if it fits my narrative.
>>105712804it's just any excuse at all to shit the thread, hell this thread had a rin op so he had to false flag then post this
>>105712573>Mikufag seething at Rin in OP, again.
>>105712804we know, just let the guy tire himself out so we can move on
>>105712804Janny baker plays both sides.
>>105712779I think the banned text is about designing women clothing, dressing a woman up and putting on her makeup. Which basically confirms that the r9k thread was the janny himself. You love to see him self report like that.
personally i like the baker, he is always reliable and i like his recaps
as far as the miku thing is concerned im neutral, dont care if its used, dont care if its not
it doesnt matter what images he use for the OP or the recap
im here for the text, not the images
>>105712885>and all that i can see, is just a yellow lemon tree
>>105712872>as far as the miku thing is concerned im neutral, dont care if its used, dont care if its not>it doesnt matter what images he use for the OP or the recap>im here for the text, not the imagesWe know you agree with what you are doing baker faggot.
For me? It's mlabonne/gemma-3-27b-it-abliterated-GGUF
i am both sides and the janny and 5 other posters in this thread and im singlehandely leading this drama from start to end
>>105712900I posted this btw
>>105712779Sometimes IP has an old ban. Are you sure it's for you? Anyway.
>>105712909Does not matter, must push Uganda.
>>105712895im not the baker you dumb fuck
this is what we mean by paranoid schizophrenia
you have no stats, no data, no evidence of anything
you can just go combine all anons into one
>>105712932>im not the baker you dumb fuckWe know you pretend not to be baker faggot.
So can we talk about local models now?
Is there any hope of Nemo/Rocinante getting significantly better replacements for roleplay?
>>105712953I'm a Mistral shill now. 3.2 is fagtastic.
>>105712804>its not me guize i swearuh oh
>>105712494Just one more year and I'll join you.
>>105712953I mean yeah, I expect that eventually there will be something better. But its impossible for any of us here to actually predict timelines beyond 2mw
>>105712969You mean the magistral 3.2 thingy?
>>105712972pic related
>>105712953Depends on your system.
For 8gb of VRAM?
Not really.
>>105712972You need to be over 18 to post on this site.
>still can't have waifu without dementia
>>105712827>just let the guy tire himself outLike how the migger got tired of spamming thread irrelevant loli ryona rape images 24/7/365 or? Oh, not like that huh.
>>105712779>You remember that image board you went to when 4chan was down?Saying what that was is the banned text. This shithole, understandably, doesn't want to bleed users by allowing anyone to know there is an alternative.
>>105712995>tfw people born in 2007 can post here
>scenario>anon unwittingly join gay wizard school>anon is the apprentice and wants to genuinely learn magic>but he is getting molested in every turn >>105712992Just the 3.2 vanilla. I mean it's good improvement over 3.1. Haven't even tried any finetunes.
the thing is when one person starts acting out, everyone else chips in and does it too just for fun
and then it spirals out of control and mods cant even bother with it
in other words, one persons trolling triggers everyone elses inner troll to start trolling
>>105713004why are you here if you hate this "shithole" so much
>>105713016is this gay rape wizard school a femboy type thing or is more of a bara sort of deal?
>>105713016Does vanilla handle loli?
How about the whole 13th century autonomy thing?
>>105713026Did I say I hate it? Maybe I like eating shit.
>>105713036That was mainly a mixtral thing, newer mistrals aren't anywhere near as bad as it was.
>>105713036>Does vanilla handle loli?
>>105713057You sure about that? I still have trauma from mixtral.
>>105713036NTA but yes it does very well indeed. I could post a log/screenshot but don't want to get put on to some special internet list even if the log wasn't seriously made - i often like to fool around for humour's sake.
>>105713026Maybe he misses old 4chan? A non-safespace one?
>>105712953Yes I'm using DeepSeek-R1-0528-UD-Q4_K_XL and I like it much more than Nemo. My settings:
min_p 0.0001 (1e-4)
temperature: 0.6 to 0.8 (at the moment I am using 0.6)
logit_bias: [ [12, -2], [666, -2], [965, -3], [982, -3], [1248, -3], [1613, -3], [2619, -3] ]
Use Onimai for the next op just to spite him.
https://www.reuters.com/world/china/deepseek-r2-launch-stalled-ceo-balks-progress-information-reports-2025-06-26/
>DeepSeek R2 launch stalled as CEO balks at progress, The Information reports
>
>June 26 (Reuters) - Chinese AI startup DeepSeek has not yet determined the timing of the release of its R2 model as CEO Liang Wenfeng is not satisfied with its performance, The Information reported on Thursday, citing two people with knowledge of the situation.
>
>R2, a successor to DeepSeek's wildly popular R1 reasoning model, was planned for release in May with goals to produce better coding and reason in languages beyond English, Reuters reported earlier this year.
>
>Over the past several months, DeepSeek's engineers have been working to refine R2 until Liang gives the green light for release, according to The Information.
>
>However, a fast adoption of R2 could be difficult due to a shortage of Nvidia server chips in China as a result of U.S. export regulations, the report said, citing employees of top Chinese cloud firms that offer DeepSeek's models to enterprise customers.
>
>A potential surge in demand for R2 would overwhelm Chinese cloud providers, who need advanced Nvidia chips to run AI models, the report said. [...]
>>105712953>So can we talk about local models now?>posts this disgusting avatar againNo. Kill yourself.
>>105713094why are they so openly lying? Why would there be a R2 without a V4?
>>105713111V4 would be a completely new architecture; R2 might "just" be an update to their RL techniques.
>>105713133V4 could be a new base pre-train, not necessarily a full arch change.
>>105713057People finally accept mixtral is bad? At the time people would get so mad at me for saying that.
file
md5: 8ba71e0cf76aaacf88ea05227fbac44c
๐
>>105713076>DeepSeek-R1-0528-UD-Q4_K_XLHow do you fit that into vram?
>>105713023But why would someone want to troll this beautiful place of friendship and tolerance?
file
md5: a65bfc82e39e02f319f9e8fb43a2c6a9
๐
>>105713191>this beautiful place of friendship and tolerance
>>105713169You don't.
It's a MoE.
>>105713171i just noticed those soldiers with guns are white
nice
what is that from?
>>105713281Somewhere in Africa idk
>>105713250>>105713302Thanks for the latest updates on what the people over on reddit are saying!
From the looks of the second screenshot, it seems like they have a twitter account where they show off clever and witty comments. I'd suggest that they also post them to 4chan, but luckily we have you already taking up the mantle to do the same!
Please stop with the reddit screens. I can only take so much cringe.
https://huggingface.co/black-forest-labs/FLUX.1-Kontext-dev
It's up.
>>105713343can it draw boob?
>>105713336You repost twitter pics here so shut the fuck up retard
>>105713377oh yeah? name 1 twitter pic i posted
>>105713377Social media is all interconnected, there's no need to be upset friend!
>>105713372>>>/vt/101618651>It's not censored like normal flux:https://litter.catbox.moe/lexhgab5akoli5tu.png
Apparently.
>>105713411no its a drawing
>>105713400>>105713419>crosslink to vtumor shitholeConsider the following
>>105713425yes asking for hug but the model don't, not very good
>>105713434Do I need to update Comfy for this? I don't want to because it has been getting progressively worse.
>>105713447Just update, it'll be fine! :)
>>105713400That's with an adaptor. Can it draw it from a raw prompt?
>>105713094Why they simply cannot obtain them by smuggling or illegally in general?
>>105713434Tell her I don't need a hug but I do need her to join the ongoing shitposting effort in /g/ /lmg/. Especially when it concerns the janny who wants to his own dressup doll.
>>105713498OpenAI And Google Hid Their Model Thinking, Costing China Months Of Time In The AI Race.
>>105713458The best part of 2025 for me so far (as far as LLM's are concerned) is me finally deleting oobabooga text webui.
>>105713094Ok so the only sentence of actual "news" in this article is:
>Chinese AI startup DeepSeek has not yet determined the timing of the release of its R2 model as CEO Liang Wenfeng is not satisfied with its performance, The Information reported on Thursday, citing two people with knowledge of the situation.and the rest is padding out word count.
I really hate to say this but, this really could have just been a tweet. Even that one sentence of content is questionably newsworthy.
Quite grim what journalism has come to.
>>105713525A silly excuse to overcharge you for tokens you can't even see
>>105713535why?
what's wrong with it?
>>105713547The Information were the ones that said the May rumor as well iirc, they're just trying to stay relevant. Nothing wrong in a little engagement farming.
Which model is best for responding to posts in 4chan /lmg/ threads?
>>105713634You don't need a model. Just post hatsune miku pictures.
>>105713650Which model was used to post that message?
no
md5: 051f6f5419d789effcc7a0f518e9ffdd
๐
>>105713667That post was definitely made by the old gpt-4chan model. Very basic.
>>105713076Oops meant DeepSeek-V3-0324 not R1-0528.
>>105713169I don't, I just have it in LPDDR5 RAM and it's usable because
>>105713227
current somewhat budget meta used rack mount server for ai?
i heard that turing, pascal, maxwell will no longer get cuda updates
>>105713717Go Volta: https://l4rz.net/building-a-poor-mans-supercomputer/
what happened to automatic1111?
he used to contribute in these threads
>Meta's AI hiring efforts have scored big: Trapit Bansal, a highly influential OpenAI researcher, has officially joined Meta's AI superintelligence team.
https://x.com/TechCrunch/status/1938310094149554531
>>105713798Dev drank himself to death. Copious amounts of vodka every day tends to do that.
>>105713798he's still here.
>>105713862Russian zoomers drop alcohol en masse btw
>According to the VCIOM poll, the share of non-drinking Russians has doubled in 20 years and reached 48%. For comparison: in 2004 only 27% of respondents consciously chose sobriety. Deputy Prime Minister of the Russian Federation Tatyana Golikova noted that the amount of alcohol consumed per person for 15 years (from 2009 to 2024) has decreased from 18 litres to 8 litres.
>>105713094let them cook, faggot ignorant of the llama4 failure
>>105713650this man gets it
When will they invent a way to train and use the big LLMs on consumer hardware?
>>105712314It annoys me too. Full automation was always my goal and if max char limit was 3k I probably would have done it already as I could post them in full unedited, but since I still check them anyway to make sure nothing interesting was cut off it hasn't been worth it.
I think first I need to have it generate longer summaries of reply chains and have the bot use those to order them against each other.
Though, I'd still make it wait until page 9.
>>105712799>Google made it clear in the blog post that the purpose of these models is to run on their Pixel phones.and yet those models are also better than the model they didn't explicitly release for phones ( 4B and 12B Gemma 3 )
don't sleep on the release just because of the phone marketing, it's an interesting model with a lot more innovations than we're used to (like the matryoshka architecture) which would benefit bigger models to run on your computer.
How often do you hear people saying they're unhappy because X model is a little above or too below their optimal VRAM point and they'd like models to be released in greater variety of sizes? it's such a common source of whine and this model architecture would fix it
https://files.catbox.moe/13ida5.jpg
>>105714003fuck that's hot
mfw this isn't my hugbox
mfw stop doing things I don't like
Whats the safe context length for mistral small 3.2?
>>105714073>safeplease include a trigger warning next time
just seeing that word makes me irrationally angry now
>>105714003>someone posts this generic anime girl in previous threads >>105704741>mikutranny sees threat in that and makes snuff porn of her
>>105713792https://www.tomshardware.com/pc-components/gpu-drivers/nvidia-starts-phasing-out-maxwell-pascal-and-volta-gpus-geforce-driver-support-status-unclear
kind of scared too
>>105714073I just set --ctx-size 32768 in llama.cpp, should use larger but why bother.
>>105714105>he didn't buy blackwell
Why I keep reading the same posts?!!!
>>105714098SNUFF PORN? WHERE??
https://www.reuters.com/world/china/deepseek-r2-launch-stalled-ceo-balks-progress-information-reports-2025-06-26/
>DeepSeek R2 launch stalled as CEO balks at progress, The Information reports
>
>June 26 (Reuters) - Chinese AI startup DeepSeek has not yet determined the timing of the release of its R2 model as CEO Liang Wenfeng is not satisfied with its performance, The Information reported on Thursday, citing two people with knowledge of the situation.
>
>R2, a successor to DeepSeek's wildly popular R1 reasoning model, was planned for release in May with goals to produce better coding and reason in languages beyond English, Reuters reported earlier this year.
>
>Over the past several months, DeepSeek's engineers have been working to refine R2 until Liang gives the green light for release, according to The Information.
>
>However, a fast adoption of R2 could be difficult due to a shortage of Nvidia server chips in China as a result of U.S. export regulations, the report said, citing employees of top Chinese cloud firms that offer DeepSeek's models to enterprise customers.
>
>A potential surge in demand for R2 would overwhelm Chinese cloud providers, who need advanced Nvidia chips to run AI models, the report said. [...]
>>105714137In your rotten axe wound of course
>>105714150why are they so openly lying? Why would there be a R2 without a V4?
>>105714098mikutranny is faking the yandere for miku thing
he posts a lot of other characters on /h/
he has a very recognizable style I've never seen that kind of eyes other than in his gens and this lora, which I believe he made and this must be his civitai account:
https://civitai.com/models/1321855/flatserious-eyes-noob-ai-xl-10-vpred
>>105714172V4 would be a completely new architecture; R2 might "just" be an update to their RL techniques.
>>105714178V4 could be a new base pre-train, not necessarily a full arch change.
file
md5: bff22620d3d14b432f430129f63d3dd6
๐
What the fuck happened to webui?
Where's the tensor cores checkbox?
Where's the no mul mat q checkbox?
Also what settings for mistral small?
the bots are looping again
>>105714178I want to know if they'll make any mention or reference to all those papers saying that RL for reasoning doesn't create new emergent capabilities, it simply reinforces existing "correct" or "best" behavior according to the reward model, meaning that it improves accuracy in comparison to the base model but not it's capability set.
>>105714201>>105714215Thanks for the latest updates on what the people over on reddit are saying!
From the looks of the second screenshot, it seems like they have a twitter account where they show off clever and witty comments. I'd suggest that they also post them to 4chan, but luckily we have you already taking up the mantle to do the same!
>>105714177you sound obsessed
go take it up with they/them if this is so important to you
>>105714150Why they simply cannot obtain them by smuggling or illegally in general?
>>105714248OpenAI And Google Hid Their Model Thinking, Costing China Months Of Time In The AI Race.
>>105714255A silly excuse to overcharge you for tokens you can't even see
>>105714150Ok so the only sentence of actual "news" in this article is:
>Chinese AI startup DeepSeek has not yet determined the timing of the release of its R2 model as CEO Liang Wenfeng is not satisfied with its performance, The Information reported on Thursday, citing two people with knowledge of the situation.and the rest is padding out word count.
I really hate to say this but, this really could have just been a tweet. Even that one sentence of content is questionably newsworthy.
Quite grim what journalism has come to.
>>105714264The Information were the ones that said the May rumor as well iirc, they're just trying to stay relevant. Nothing wrong in a little engagement farming.
>>105714210They removed a lot of things.
>>105714150let them cook, faggot ignorant of the llama4 failure
>>105714280Why? Do I need to revert? What the fuck is going on?
>>105714264>Quite grim what journalism has come to.anon, if you could get a job like this where you don't have to do actual work and you would get paid anyway
wouldn't you jump for this opportunity? I would
I don't hate those journalists, I would do the same if I could
>>105712608>Gemma 3nIt's been out for quite some time in preview. I guess this is them doing some final tweaks to release it. It is really competent though, this is better than what used to be 30B from what I have tried, and my phone doesn't get hot from using it.
>>105714307Dunno. I guess they're going the service tesnor route.
file
md5: 2b760e25fa8fab43b0a5f0d43e0ab91b
๐
>>105714177>tranny>yurifag Checks out
And which mistral template for the mistral small 3.2 model?
Mistral V7?
>>105714405Mistral V3 Tekken works the best for me. There was a debate about this few threads back.
>>105714430Thanks, I'll try with my modified v3 tekken template then which I use for rocinante.
How about temperature? 0.15 like recommended by mistral or is it 0.8 like nemo?
>>105714432I don't have anything above V7 in my sillytavern.
>>105714457https://litter.catbox.moe/fv1vxotzlnal5yss.json
Here's a template from the old thread.
It's been great at least for me.
>>105714457>I don't have anything above V7 in my sillytavern.>https://github.com/mistralai/mistral-common/blob/535b4d0a0fc94674ea17db6cf8dc2079b81cbcfa/src/mistral_common/tokens/tokenizers/instruct.py#L778Seems the only difference from V7 is tool calling.
https://x.com/GoogleDeepMind/status/1938278533517746686
>>105712466Looks just just downloading strait from hf, its a nice format since everyone uses the same format
file
md5: 02353177101cd97210a36d8383ec4711
๐
>>1057144670.6 temp then.
It seems kind of uhh... intense.
My rocinante system prompt might be too much for it. I've been perfecting my system prompt for rocinante for nearly a year now...
Now I have to start all over again. Is mistral small 3.2 really worth it?
>>105714177bwo really needed to make a lora for jitome?
Does silly tavern not support dropping in images for modal models to describe?
>>105714617You have to send it as an attachment I think and it only works using the chat completion API, maybe.
file
md5: b229a057cadeed4f2867691e1f69dcd6
๐
Doesn't this mean it can describe images?
>>105714264>R2 is not out yet because they haven't finished R2 yet>that will be $750 a yearI sure love The Information's very Credible Sourcesโข!
This is much better than I hoped. It can track simple values from pretty far away or maybe this was just luck I don't know.
>>105714676Just a multi-turn conversation with some fixed initial prompt or something more involved?
>>105714660It seems like it only works the first time you send and not if you regenerate the reply.
file
md5: 6e427c6bdde6856ace64f1835ab89b3b
๐
>>105714660Switching to chat completion made it schizo, and it definitely does not work.
>>105714760What backend are you using?
Did you try the "Send inline images" option?
>>105714742Testing the waters how much it can remember so this is just multi-turn conversation.
I am going to implement tracked variables as soon as I get my bearings.
I'm just surprised how well it stays together because in general. Of course this is all relative, to someone else this is probably shit.
There was one interesting example card with variables and stuff here:
>https://chub.ai/characters/horcocks/uss-galaxy
>>105714771ooga webui
Where is that option?
>>105714801Should be in the same panel as the samplers when using the chat completion API I think.
Also
>ooga webuiRight, but which backend? Llama.cpp, exl2, transformers?
Have you tried using the backend directly without ooga in the way?
>>105714822llama.cpp, but ooga webui provides the backend api, not llama.cpp
file
md5: 24aa1471e2f306bff52d24e8bdfb5f28
๐
>>105714822Also there is no such option. Do I really need to abandon ooga to get images working?
>>105714887>Do I really need to abandon ooga to get images working?That I don't know, but trying with llama-server would be the next thing I'd try in your place.
Raycast on windows is cool
file
md5: cf5bec590f4e775e31113faa4876c183
๐
/b/ here
QRD on this Miku thing?
>>105714912I decided to not care about sending it images instead.
Now how do I get mistral small to stop abusing emdash?
>>105714993Ban the token I guess.
>>105714993Ban the token.
>>105714993>>105714996Oh yeah, about the image thing. I don't know if ooba does that automatically or whatever, but you did load the mmproj so that the model could see the images, right?
Does non-english training make the model smarter overall/in english?
>>105715011Load the what? Didn't see any such option. I'm assuming ooga handles whatever that is.
>>105714996>>105714997Didn't help at all. Banned tokens never did much to begin with.
It managed to use emdash 6 times in one response after "banning" the token.
>>105715066Check your token probabilities, you're doing something wrong.
>>105714987wow. that answers absolutely nothing and just looks like incoherent schizobabble.
Wtf it's Miku. Who gets all autistic over fucking Miku?
>>105715075My token probabilities are 100%.
>>105715085It looks normal if you think for a second.
>>105715066>Didn't help at allDid you check if all the emdashes are the same token probs?
>>105715049>Load the what?Yeah, you should look at ooga's docs to know if it loads that automatically.]Multimodal models with llama.cpp are separated into the normal llm (the text part) and the image part (the mmproj) that you have to load alongside the text part.
>>105715094Miku is fine, the faggot behind her is not.
>>105715085The guy who replied to you gets upset by any and all images of vocaloids and most of his posts are incoherent schizobabble.
>>105715096my country revamped its keyboard layout standard and added the emdash as one of our keysโnow I'm tempted to use it all the time to pass off as an AI.
>>105715075Token probabilities has never worked for me in silly tavern, even with the option in user settings ticked.
>>105715129It needs to be supported by the backend.
>>105715111you sound obsessed
go take it up with they/them if this is so important to you
>>105715108I still don't quite understand.
>>105715111>the migger is still coping and defaulting to calling others the schizosgrim, the chuds really broke this low iq retards brain yesterday
>>105715123Which country?
>>105715167https://learn.microsoft.com/en-us/globalization/keyboards/azerty_english
>>105715097no, the first one very much reads like someone who has no idea how the fuck 4chan works and blaming it on mod conspiracy theories because they're too retarded to understand that you can't report text posts as NSFW site-wide.
I also have no idea what the fuck the r9k shit has to do with anything. It's just textbook schizobabble where you try and make one or two sort of plausible comments(that are bullshit to begin with) and then bring in something completely unrelated and act like it's 100% verified totally true FACT.
>>105715123I'm from scandinavia and the layout is so fucked up that I just always use US layout because it's the easiest to type with especially when scripting something.
I still can't fathom who is the retard who messed up everything else just because of รค/รถ letters holy jesus christ. They basically switched everything else around too.
>zoom out, make her breasts gigantic, her hair very long, looking at viewer, split her hair color at her templer so the left side of her hair is gray
Mostly censored and a little limited in what it knows but pretty good, takes some rerolling for more complex things but what doesn't
Shows great promise that every model in the future, one that is not censored and capable of proper realism like Chroma, will one day have this easy to use textual editing workflow
Fuck this mistral small shit. I'm going back to rocinante.
The emdash has mkultra'ed me into wanting to murder.
>>105715149I think you do.
https://desuarchive.org/g/thread/104414999/#q104418525
https://desuarchive.org/g/thread/104414999/#q104418574
Mikufag is the thread baker and janny, he often samefags to hell for optics and throws ban at anyone he does not like. Needless to say he spams his favorite waifu all day 25/8 non-stop making this thread another /sdg/ or /ldg/ cesspit like these /v/ sofcore porn threads.
janny list from that leak https://web.archive.org/web/20250617190717/https://rentry.co/o84vftsb
>>105715218have you tried finding a post with an image attached and trying to report it to see the dropdown, retard?
>>105715176Interesting, but on Linux you can easily type them regardless of the keyboard layout with compose key combinations.
https://en.wikipedia.org/wiki/Compose_key#Common_compose_combinations
>>105715176Are em-dashes more commonly used in french?
>>105715218>hiro - [email protected] - adminHiro is my sensei.
>>105715213Meant to post about https://huggingface.co/black-forest-labs/FLUX.1-Kontext-dev on /ldg/
>>105715276Damn, I thought it was a finetune but it's another official cucked and distilled model.
>>105715245not in every day speech (and particularly not before being introduced in our keyboard, normies wouldn't learn the alt combos)
but it's a pretty common punctuation for writing dialogues in fiction instead of quoting with "" or ยซยป you do it like this
โ Hey, how are you today?
โ Fine, and you?
when you ____ away
you don't ____ me say
please
oh baby
don't __
>>105715334How do you feel about Kingdom Hearts being less and less final fantasy adjacent?
What the fuck do I need to do different to see token probabilities in silly tavern with llama.cpp?
llama-server --port 5000 --threads 32 --ctx-size 30720 --flash-attn --gpu-layers 41 --cache-type-k q8_0 --cache-type-v q8_0 --model user_data/models/Mistral-Small-3.2-24B-Instruct-2506-Q6_K.gguf
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
MOTHER
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
FUCKING
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
EMDASH
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
>>105715461Nothing as far as I know.
It should just work.
>>105715185You wanna see real schizobabble? Here you go --->
>>105661688
>>105715566I knew what you were going to link before checking.
>>105715185>conspiracy theoriescia buzzword
>because they're too retarded to understand that you can't report text posts as NSFW site-wide.Kek, tranny gets caught for the second time lying trying to revise history after already being proven wrong. the reported post HAD A CROPPED PORN IMAGE, disgusting troon.
https://desuarchive.org/g/thread/104414999/#104416649
>>105715604but thats a girl
>>105714098That is some dedication to the mental illness...
file
md5: 3ade7439ca20c589cfd0fc84fe79e58f
๐
>>105715594Why would i link without checking it?
Its literally nothing but image slop from one tranny spammer, the thread serves as containment zone only, nothing else of any value.
>>105715612Fhanks for finally conceeding that you are a bad faith digusting tranny.
Cropped porn is against the rules. Try posting it as a new thread on a blue board with a relevant topic and see how fast you get banned, AGP fag.
so bros, when did your summer break start?
started a few weeks ago for me :)
>>105715641B-But... xhe wants to goon with xhis sisters in xhis favorite general!1!
>>105715604and the archive says that that post specifically was reported? and then go on and pretend that other posts in the thread suddenly can't be reported for NSFW(that's not something that can be disabled on a per thread basis).
take your medication or at least take a hammer to your walls to let whatever is crawling in them out.
>>105715631>Why would i link without checking it???
my brother in christ a 4b language model has better reading comprehension than you do
it means it's time to go to sleep for you and if you're in a timezone where should be well awake and alert I would consider brain surgery
flux kontext dev is very cool
>>105715185>mod conspiracy theoriesEat shit and die faggot. Bans in this thread are absolute bullshit while mikuspam is never moderated. It is not a conspiracy theory. This is a troon hugbox and you are part of it. Kill yourself.
>>105715641>Cropped porn is against the rulesSpoilered porn is against the rules, ESL-kun.
file
md5: de190d0a5150bef6025e043f0e8fd313
๐
>>105715218Now i wonder which one of these is our janny
>>105715631Dipsy is cute tho
>>105715641been to /vtai/ recently dipshit
>>105715664>and the archive says that that post specifically was reported?agp tranny is so braindead from his hrt that he cant follow basic conversations
yes, retard, that is the post that is quoted https://desuarchive.org/g/thread/104414999/#104418525
>and then go on and pretend that other posts in the thread suddenly can't be reported for NSFWagain with the clockwork tranny tactics of gaslighting, despite proof posted in the same comment posted with the allegation the https://desuarchive.org/g/thread/104414999/#104418574
You don't understand guys. Sam Altman is reading this thread and laughing so hard when you dance like puppets for him. We need to be united! We can't let Sam win. And Sam doesn't win when we post more Miku.
>>105715631>"1girl, portrait, standing, front view" boredomWow
>>105715731>proof>click report on a text post>whoa i can't report it for nsfw wtf troon mods jannies are conspiring against me troon troon troon
>>105715731>also, i reported the post for nsfw and after it didnt got banned, of course, i now see that you cant report any posts in this thread for nsfw at all? didnt know jannies can disable it thread wide so easily>also, i reported the post for nsfwdo you have the reading comprehension of a snail
also the crop doesn't show what they were trying to report. Did you know that you get a completely different report dropdown when you try and report a post with an image vs one that is just a text post?
Now you do! :)
>>105715742worthless, shits the thread up, incentivizes further shitting, ensures that shit wars will never stop
>>105715761>troon troon troonthis but unironically
>>105715762none of this shit matters. it is offtopic spam. and janny allows it. so have fun having your thread shitposted to oblivion.
>>105715684I push The Blonde into the saltwater. Sand goes into her eyes. Then I run away while urinating as my cock flops up and down spraying everywhere. She cannot see this unfolding.
>>105715761Yeah they literally are, that's how these men enforce rules and then ban you for whatever reason.
>>105715797>posts actual trash
file
md5: ffe2bacb14ea67d63a4d4982d408e6e5
๐
>>105715594Why would i link without checking it?
Its literally nothing but image slop from one tranny spammer, the thread serves as containment zone only, nothing else of any value.
>>105715761Yeah they literally are, that's how these men enforce rules and then ban you for whatever reason.
>>105715797>>105715815>>105715797>>105715815xhes back after his dilation slash gooning session
file
md5: 1b4a6aeff72b317b5e4c1e8b0289bc3f
๐
GOD DAMN IT WHAT IS WRONG WITH THIS SHIT WHY WONT IT WORK
>>105715880It just works in mikupad
>>105715762the post was made in the middle of the thread when anyone could have validated the claim, retard
>passive aggresive cope emoji after being btfod for the 100th time todayno matter the cope, you will simply never be a woman and forever be an agp retarded migger, kek
im glad people are absolutely clowning you in every thread the last few days, and the only thing you tried to do every time is try to gaslight to play it off hahahahahah, grim.
>>105715880When you request the log probs, there's a thing that shows on llama.cpp's console. See if there's anything there.
>>105715880Works on my machine with llama.cpp server
>>105715886>hahahahahahdude you pwned them so hard
>>105715886Funnily enough i think he got that extension replacing "tranny" to "you are valid" or whatever snippets.
ive been over this with sdg avatarfags vs anti-avatarfags and hdg highlightfag vs anti-highlightfags
the only solution is to split the general because neither side will back down
file
md5: d1a5fc7ec05211377296b005f2502771
๐
>>105715892There's nothing.
>>105715916see you on in every other Local Migu General
or Hatsune Designated General, aiming for 1000 threads.
>>105715916honestly theres so little to talk about maybe its time to question if local models need a general in the first place
>>105715916He'll get tired eventually.
>>105715932/wait/ one definitely needs to die for good, /lmg/ draws near its death too.
Btw the OP mikuposter is posting porn in /ldg/ in direct attached image that I reported shortly after it was posted and its still up
>>105715769Seems like everyone is right about him being a janitor/mod.
>>105715886>the post was made in the middle of the thread when anyone could have validated the claim, retardand yet no one corroborated it, either.
Also their cap happens to be missing every image specific report option. Almost as if they weren't reporting a post with an image attached.
How queer!
>>105715916having one general for local models and one for culture war obsessed tech illiterate election tourists sounds like a good idea
we've been over this before and even when he gets his own thread he just shits up the regular thread anyway
the only solution is to not feed the trolls
report hide and ignore
>>105715949I reported your gore and it still up, you must be a troon janny.
>>105715916>>105715932yes please!!! death to lmg!! finally
>>105715949>getting upset about no nipples or genitalsSo the recent spam is by one of the redditors who ended up here when their local model thing died? Color me surprised.
>>105715949>attached image that I reported shortly after it was posted and its still upAw does snowflake suddenly need the jannies to protect him from the stuff he doesn't like?
>>105715949Same image = same person it's the only possibility
imagine being so shit at editing that not evne kontext wants to help you
>>105715962that post has the option to report for NSFW though?
Doesn't that completely shatter your false conspiratorial illusions?
>>105715971>incoherent schizobabble powered by whataboutismColor me surprised.
>>105715922Interesting.
You did tick the "Request token probabilities" checkbox in Silly, correct?
>>105715970Why do you come here if you don't want to be here?
Just fuck off already. Some of us likes this general and actually discuss shit.
Personally I like helping people.
>>105715963>no ukek
>>105715971>>105715972least obvious samefag that types in the same way while both defaulting to the most low iq response of a libtard tranny calling someone a snowflake while he breaks the rules while reporting others and removing other comments against him because "they brake the rules"
him posting porn is ok but people calling him a nigger he bans, kek
>>105715949Told ya, again.
He made it his whole personality, now he can go fuck himself on appropriate boards - /trash/ or /b/.
160s a gen for 20 steps, you leaked more info, shit GPU you got there. You're on a roll today with showing your hand, keep going.
>>105715461Logit bias doesn't seem to work for me either with Llama.cpp / SillyTavern.
If I were a local in this general I'd spam Miku pictures
file
md5: d08c9a374fdc298c31c70282d12718b6
๐
>>105715987Yes of course I did. It does not work no matter what I do.
I even used the latest b5760 cuda release of llama.cpp.
What else am I missing?
>people shit on AGP tranny mikuposter all day
>only he posts back at those people
>someone exposes the AGP tranny, again
>suddenly 5 comments within 1 minute in his defense appear to derail and damage control, all of which dont even have an argument
man... this is just sad.
>>105716007I have no idea.
I'm so curious I'm closing palworld to test it.
>>105716008Of course he would be in Discord, that's a given for him.
>>105715997That's not the guy you're obsessed about then. He has a 6000 blackwell.
>>105715949that isn't even the same guy. The person you're schizo'ing over uses forged or reforged and obsessively inpaints.
The person you're crossposting is just some retard trying random comfy shit using a random image they found.
>>105716008isn't it inspiring how many people actually like migu vs your entirely manufactured 1-fag show
it's not even orchestrated or arranged, that many people just dislike you entirely organically
>>105716008>in his defense
>>105715989>Personally I like helping people.C-can I dress you in a skirt and put makeup on your face? It would help me a lot.
>>105716020And my dad works at OpenAI, now what?
>>105716037No. I always get hard wearing a skirt then it sticks out and looks weird
>>105716027>random image they foundkek
ywn
baw
>>105716042>jealousy kicking inHe posted having one and he made wan loras using it.
Many other anons one.
I have one.
>>105715972Either none of it is ok or everything is ok. Pretty simple.
>summer vacation only started this week
Only 2 and half more months of this left to go.
>>105716042the person you're schizoing over is an actual person who's recognizable and has existed on another board for years.
>>105716052>appeal to jealousy Non-argument.
>>105716058guy's unemployed don't get optimistic now
>>105716061vs appeal to tranny yeah
>>105716060>don't you dare touch my favorite e-celeb!Cool
>>105716013Token probabilities only work with the STREAMING option set. I have always hated that option so it's always been off.
Now that I can see token probabilities, how does it help me permanently murder emdash from mistral small?
>>105716058This is just the schizo reheating the thing when he cracked cudadev's trip to post blacked miku. He'll get bored of it in a week when it's clear that nobody is buying his shit.
Excuse me I have an important announcement *clears throat*.
Thread culture.
>>105716061I made my argument here
>>105716020 and the only thing you did was express your poorfag jealousy here
>>105716042i had no other argument to make. I was just making fun of you for being poor.
>>105716048He makes double-replies now lol
file
md5: 23bd2a08fa05171ed07ea3cfb333112d
๐
>>105716068would rather be a known attention whore than an unknown one
anyway you are known, go back to discord
>>105716085This is why I logit-bias against em-dashes ever since I began using v3-0324.
>>105716028>isn't it inspiring how many people actually like miguno. kill yourself
file
md5: a7c0eed11958ff42a7de95016c5de7ca
๐
>>105716092I already tried that too. Emdash is 1040. How negative do I have to set it to completely remove it?
>>105716069>Token probabilities only work with the STREAMING option setJust tested and it worked with and without streaming.
Anyhow, now you find the token ID for the emdash and ban that.
With luck, it's its own individual token rather than being various tokens like a dash + the start of a word or something.
file
md5: 91e412abdd6f071875204afddb4b9bf0
๐
Kontext warm run (models loaded) vs cold run (no models loaded) @ 1024x1024, 20 steps on RTX 6000
>>105716073>reddit cultureLeads to picrel eventually.
I too can call any spam - "culture", it wont make it any better though.
>>105716047>most honest mikutroon
>>105716028Many people got over with her years ago.
The recent hype she got is entirely centered around twitter trannies
>>105715815 (filename lol) and zoomers playing Fortnite slop.
>>105716068>e-celebI really don't care. I'm just trying to tip a cow that is so fucking deluded they think some extremely random(but recognizable) person is a powertripping mod for no real reason.
also you did find their civit account but they have a twitter too @brittle_404
you sure do loads of research on miku, black men and trans people anon!
that's great you do so much activism!
you go fellow human!
In fact you go out of your way to collect and collate the info, talk about it all day every day, save screenshots and make illogical inferences constantly about everything you claim to hate!
that's totally owning them and showing how much you don't like it, spending 100% of your time absorbed in it!
file
md5: c0d4f92f2d4338451fa54598a4f0d4de
๐
>>105716133>but they have a twitter too @brittle_404What a nice
>>105716123 timing congratz
>>105716137I don't even care for your drama with others but the only fallacy machine ITT is you with ad hominems and appeal to triviality fallacies.
>>105716103It's the token counter, not the token probabilities that show the token IDs. Also the IDs seem to change... first it was 1040 now its 1674 for emdash. What the fuck?
What negative value do I even set for the logit bias?
How come banning the token has zero effect?
I'm about to declare the entire fucking model a failure for roleplay.
>>105716153He is clearly seething now
q-quick, chatgpt how do I make it feel like I'm still making sense
>>105716072Fun fact: I am the blacked miku poster and I didn't crack the trip. Also I am c.......
>>105716158Rocinante is superior for the lack of emdash alone.
>>105716137>>105716164lol stupid nigger outed himself bigly
go die for it
>>105716164peak effeminate retort. keep at it and some people may think you pass.
>>105716158>What negative value do I even set for the logit bias?I'd just ban the token.
>How come banning the token has zero effect?it should.
Try banning some really common token like " the" or something like that. See if that does something.
>Also the IDs seem to change... first it was 1040 now its 1674 for emdash. What the fuck?Look at the model's tokenizer in the original repo with the safetensors.
See if those IDs make sense. Could be that the token counter is showing the wrong token IDs for whatever reason.
>>105716007Nothing wrong. There are many tokens containing em-dashes in the Mistral Small 3 tokenizer. Try banning these:
[1674,2251,2355,18219,20202,21559,23593,24246,28925,29450,30581,31148,36875,39443,41370,42545,43485,45965,46255,48371,50087,54386,58955,59642,61474,62708,66395,66912,69961,74232,75334,81127,86932,87458,88449,88784,89596,92192,92548,93263,102521,103248,103699,105537,105838,106416,106650,107827,114739,125665,126144,131676,132461,136837,136983,137248,137593,137689,140350]
>>105716210Logit bias in chat completion doesn't seem to work, by the way. Text completion token banning works. At least as of the latest llama.cpp.
>>105716158>I'm about to declare the entire fucking model a failure for roleplay.Skill+brain issue. It's not as bad as you're making it. And you can also add in the instructions "Do not write em-dashes (โ)."
>>105716210Thank you anon those token IDs match many of the ones I've been finding now.
Also adding that full list to banned tokens completely removed the emdash from outputs, although it tries to use regular dash instead now... but at least it's not as bad.
I'll now proceed to genuinely test the model to see if it has any value for roleplay at all.
>>105716247LMAO no. Instructions telling a model not to do something do not work at all. ESPECIALLY for mistral based models.
so, is this guy just angry that his janny application got denied again?
>>105716247>"Do not write em-dashes (โ)."Small models do not handle negatives well, and it probably will skirt around it through technicalities since there are other tokens with em-dashes
>>105716210 where it can comply but still piss you off.
>>105716133>Look it up>Its all loli slopCreatively bankrupt.
>A delicious tremor passed through him at her reaction.
Dear fucking lord it's new shivers.
>>105716247That does not work. You can beg Deepseek to not do em-dashes like that and it'll do them anyway.
>>105716106That picture is just programming culture.
Fortunately I decided to major in something else.
Is he a troon? Why yes he is and that is the issue.
>>105716286Many people obsessing over anime characters troon-out at some point in time, its like a fixed constant at this point hence the zoomers with "tranime" catchphrase, a counter-culture.
Is this recurring episode part of thread culture?
This post is art.
This post is thread culture.
etc.
Mistral small 3.2 needs these two extra stopping strings or it does funky empty lines and large text formatting.
>"\n---\n", "**["
>>105716284It works in Mistral Small 3.2, though, at least for simple examples.
Gemma sirs...
>some Gemma derivative models (maybe Amoral and Fallen, but definitely Tiger-Gemma, Big-Tiger-Gemma, and the abliterated models) are in violation of the license, and it might be within Google's legal power to tell Huggingface to delete the repos for such models
https://www.reddit.com/r/LocalLLaMA/comments/1llcyvu/lets_talk_about_googles_gemma_license/
>>105716274>look it upyou were looking for a single image that was gay to pretend it was evidence but really to crank your pecker to
>>105716373Sorry Drummer, Google's legal team will be contacting you shortly.
>**, **", "**
[1438, 55387, 58987]
Ban these logits too to get rid of bold text abuse...
Frankly at this point a fine tune of mistral small 3.2 would be better.
>>105712608impressive, what a good little RAG slave.
i hope it also works well for MCP servers
>>105716371>projecting pedotard goes "yuh ghey for not approving my retarded shite!1!"k
>>105716434>look it up>it's all>allyou keep playing your hand
>>105716414If it abuses bold text it's probably because you have that in your card/system prompt. The previous Mistral 3.1 also behaved similarly. It occurred with italicized text too.
>a fine tune of mistral small 3.2 would be better.Never.
>>105716440Go post that in your Discord if you want universal approval, nigger
>>105716449oh which discord would that be
>>105716366Prolly because all these are used for pedoslop and joogle does not want bad publicity.
I don't understand the logic behind posting gore on 4chan.
It has no shock value here. All it does it get you banned and then you have to change your ip or whatever.
>>105716461>this isn't your hugbox>sees a migu>chimps out
>>105716461>multiple newnigger retard implications
>>105716461it's his fetish, just like bbc and scat
>>105716461Dead niggers are funny.
>>105716445No, no such thing in the card or system prompt. I despise asterisk use in my roleplays, so I always strip them out of all character cards. Rocinante/Nemo didn't have this problem either.
I always prefer fine tunes over mistral base models.
I hope someone gets to making some nice meme merges of this model. I noticed it released just 6 days ago, so it's still very new.
>>105716461saving the thread (by posting dead niggers in a spastic fit)
>me when I post organically
>often in threes.
>me when I imply I'm on the same discord server(s)
I can't stop playing my hand
>>105716461It makes redditors die inside a little every time they see those. I think he deserves a medal for making the sacrifice.
>>105716471Three shots - all missed.
>>105715990>this guy got banned quickly for btfoing a tranny without breaking any rules>>105716106>this guy got banned quickly for posting a compilation of trannies on twitter>>105715850>mikunigger posting direct attachments of porn is left up for more than an hour until people complain that its too obvious that they reported him and its still up>>105715949>once this is mentioned you have 5 comments seething in the replies within 1 minute with no argumentsagp mikutroons really are a whole circus lol
>backlink to a dead post
>claims it's still up
what is this retards game
>>105716529>is left up for more than an hourHe implied it was up for hours there you stupid nigger
>>105716520it's literally all you
>>105716535do you happen to know when it was deleted?
oh. wait. it was deleted with all the other trash that was deleted in this thread. Simultaneously.
Almost as if no one was even watching the queue until then. Curious!
Mistral Small 3.2 definitely disobeys the system prompt and user responses much worse than Rocinante.
>Explicitly cut off the hand of a character.
>AI decides to ignore you and say that the blade stopped just in time before harm was done
lol
>>105716520>this guy>this guynobody believes that you are more than one person, bbc scat lover.
>jannies delete gore one by one after leaving it up for hours
>do absolutely nothing about about the spamming, flaming, and trolling
why?
>>105716545>appeal to lack of awareness on what happens in other threads
>schizos out over bullshit boogiemen in the shadows
>gets banned
>ban evades
>samefags himself as if there's some sort of proof there
>wonders why everyone is calling him retarded
>>105716554>the guy shitting on the only people who post bbcope aside from jews (which are trannies) is actually the tranny!accuse your enemy of what you are guilty of indeed, thanks for admitting that you are also into that, tranny.
did an entire mental hospital suddenly get into LLMs?
>jannie mikutroon deletes comments against him within minutes
>his porn isnt deleted for almost an hour and a half
>he says that people complaining are actually "just one guy and a snowflake who is also into bbc :)"
Kek, you really can't make this shit up
file
md5: 7608ef6c57f755005330ea9ff562c9b9
๐
have fun malding
>>105716573No, it's just Sam Altman trying out his new shitposting bot.
>>105716573>did an entire mental hospital suddenly get into LLMs?>suddenlythe troon mental hospital has been here
This guy sees Altman everywhere he goes
>>105716590 and is totally sane btw
>>105716573one guy, almost TWO years
>>97811125
>>105716608I can't believe how much shitflinging I caused by writing that Miku prompt shell script for the first leaked llama, I'm kekking through every thread
>>105716483I also approve giving that guy a medal.
I would like to close this discussion by bringing up key evidence everyone ignores. I remind you tgat cudadev gas endorsed mikuposting. That is it.
He also endorsed hitting that feminnine jart bussy a bit later on but never mind that.
>>105716573it's a rerun, we've had this episode before
same meltdown about miku and troons, some distasteful spam, some falseflagging, muh jannies, all the thread regulars are trans and jewish, etc etc
little tyke will tire himself out sooner or later
>>105716637>He also endorsed hitting that feminnine jart bussy a bit later on but never mind that.Exactly, a literal tranny enabler endorsing the tranny OP, lol.
>>105716638>distasteful spamAll of it will go away once you decide to log back in. That place even works again after that mod left.
>>105716638You know you are definitely not a redditor or discord tranny when you totally randomly defend jews and trannies on 4chan, anonymously, for free. Kek.
>>105716637>He also endorsed hitting that feminnine jart bussy a bit later on but never mind that.He kinda has a point mikusisters... Whenever you look closer there is always a troon at the end of the mikuspam trail.
I'm endlessly swiping with Mistral Small 3.2. It seems incapable of producing mistake free responses.
There is always some kind of problem that makes it unacceptable. It's not as fast as nemo eithe which makes it even worse.
>>105716735^
one of the only person not here to astroturf mistral
dogshit models man
>>105716665'anime spammer to tranny' pipeline is real.