Thread 106522347

/lmg/ - a general dedicated to the discussion and development of local language models.

Previous threads: >>106516368 & >>106512307

►News
>(09/05) Klear-46B-A2.5B released: https://hf.co/collections/Kwai-Klear/klear10-68ba61398a0a4eb392ec6ab1
>(09/04) Kimi K2 update for agentic coding and 256K context: https://hf.co/moonshotai/Kimi-K2-Instruct-0905
>(09/04) Tencent's HunyuanWorld-Voyager for virtual world generation: https://hf.co/tencent/HunyuanWorld-Voyager
>(09/04) Google released a Gemma embedding model: https://hf.co/google/embeddinggemma-300m
>(09/04) Chatterbox added better multilingual support: https://hf.co/ResembleAI/chatterbox

►News Archive: https://rentry.org/lmg-news-archive
►Glossary: https://rentry.org/lmg-glossary
►Links: https://rentry.org/LocalModelsLinks
►Official /lmg/ card: https://files.catbox.moe/cbclyf.png

►Getting Started
https://rentry.org/lmg-lazy-getting-started-guide
https://rentry.org/lmg-build-guides
https://rentry.org/IsolatedLinuxWebService
https://rentry.org/recommended-models
https://rentry.org/samplers

►Further Learning
https://rentry.org/machine-learning-roadmap
https://rentry.org/llm-training
https://rentry.org/LocalModelsPapers

►Benchmarks
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
GPUs: https://github.com/XiongjieDai/GPU-Benchmarks-on-LLM-Inference

►Tools
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

►Text 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

►Recent Highlights from the Previous Thread: >>106516368

--Papers:
>106518049 >106518160 >106518284
--Chemistry-focused AI models and their implementation challenges:
>106519688 >106519719 >106519768 >106519776 >106519866
--Vibe voice synthesis quality optimization through sampling parameter adjustments:
>106519850 >106519865 >106519873 >106519909 >106519945
--Minimum dataset requirements for training basic LLM intelligence:
>106516402 >106516520 >106516631 >106516716 >106516994 >106517256 >106520214 >106520372 >106520566 >106520931 >106521020 >106521093 >106521245 >106521342 >106521445 >106521718 >106516718 >106516764 >106516846 >106516881 >106516891 >106516849
--Alternatives to HuggingFace for sharing voice model datasets:
>106517886 >106517950 >106517974 >106518002 >106518054 >106518269 >106518744 >106519046 >106519102 >106519140 >106519173 >106519692 >106520056 >106519150 >106518003
--Optimizing voice synthesis model performance on constrained hardware:
>106519713 >106519721 >106519722 >106519742 >106519774 >106519767 >106519782 >106519837 >106520969 >106520983 >106521033 >106521284 >106521368 >106521669
--Layer skip technique for speed optimization in neural networks:
>106518498 >106518519 >106518527
--Analyzing AI model performance benchmarks:
>106517499 >106517551 >106517582 >106517591 >106517629 >106517557 >106517593 >106517602 >106517638 >106517912 >106517680
--Optimizing GLM 4.5 quantized model performance on dual 3090 GPUs:
>106517197 >106517206 >106517214 >106517226 >106517228 >106517242 >106517303 >106517350 >106517359
--Step count optimization improves TTS consistency and quality:
>106516529 >106516806 >106516823 >106518407 >106519565
--llama.cpp macOS Metal backend optimization and async improvements:
>106521190
--VV voice synthesis for visual novel game development:
>106520293
--Miku (free space):


►Recent Highlight Posts from the Previous Thread: >>106516369

Why?: >>102478518
Enable Links: https://rentry.org/lmg-recap-script

>a general dedicated to the discussion and development of local language models.
>and development
What is (You)r excuse, Anon?

How do AI Agents work?

>>106521735
never forget kids
4bit quants of anything is a cope
and lower than 4bit is an even worse and more retarded cope

>>106522345
Are you retarded?
>https://char-archive.evulid.cc/#/definition?source=chub&type=character&path=mr.blackwood%252Fbelle-delphine
Atrocious.
And this one... chatgpt world salad slop.
>https://chub.ai/characters/imperatorcreator12/Lopunny
Please stop giving advice. Or maybe you are thinking you're pretty intelligent...

>>106522409
The first one is so fucking funny. It's straight out of 2020 AI Dungeon

>>106522409
in other words, you love these cards!
i knew you'd like the belle delphine one :)

woah

>try to talk lewd to llm in normal chat
>it refuses
>hook it in through silly tavern
>it unhinged
What does silly tavern do to them to cause this? Is it all just the system prompt?

>>106522373
Who?

>>106522540
You just tarnished my girl. Asking for a friend, what did you use to make that?

>>106522535
My experience is that if you use the wrong instruct template it makes the model retarded, frequently too retarded to let it remember that it's supposed to refuse.

reminder >>106519850

>>106522345
Cheers. I use chub when I can but not many vanilla or just plain vanilla sfw characters on there. That directory site is sweet, thanks for the link

>>106522373
They do it for free
>capcha: arara

>>106522761
very welcome. have fun.

>>106522535
it depends but sometimes models actually are inclined to act according to the prompt you make it answer

>>106519850
https://voca.ro/137n0Zb02BUZ

>>106522913
she sounds like a cum goddess
pls post the ref audio

Newb question. I know I'm running locally but are there any models or backends I have to worry about that are sending telemetric data to a company? Do any models perform internet searches or is it all off training parameters and my tuning/supplementary cards 100%

>>106522913
Nice.

>>106522928
If you buy a copy of Slay the Princess, all the voice files are packed up into nice little isolated audio wav files in the game directory. You can just take a handful of them, throw them in Audacity and have it arrange the tracks end to end.

Trust me, you'll appreciate the voices more if you spend some time with them first.

https://voca.ro/1oMPr4RzJTgg

>>106522993
ZAMN that's from slay the princess?
worth the pirate just for me to grab all those files then. thanks.

>>106522931
I don't know about any shenanigans with backends, it wouldn't by an interesting target anyway.
Frontends are another story, local shizo says ComfyUI pings google now and then.
Internet search and other MCP/RAG/Agent shit is something you have to set up yourself, you won't be getting it by accident.

>>106522993
What is the max audio file duration one should use?

>>106522993
>wav files
Why.

>>106523040
lossless and uncompressed

It's up
https://huggingface.co/IndexTeam/IndexTTS-2

>>106523054
ZAMM
TTS Chads eating good this month

Wtf, somehow even though I've looked at the github like at least 10 times by now, I just saw that nolima was made by adobe research. Somehow I just never looked at the URL or the title of the page.

>>106523054
SOTA just dropped
https://index-tts.github.io/

https://voca.ro/168YeYwGpMRL
Okay, deterioration over time is definitely worse at high step values.At least, with a high cfg.

>>106522928
https://www.youtube.com/watch?v=aI1-GRisd1k

>>106522965
pic related

>>106523037
I haven't figured that out yet. Sometimes I feel like I get better results with a really long (6+ minute) sample, and sometimes I feel like I do better with something cut down to one or two.

>>106523054
fuck

>>106523054
>0 training code
DOA

>>106523054
Can it do sex?

>>106523049
Steam should start charging gamedevs for storage. We have some insanely good speech audio codecs, and these fuckers put wavs into their games.

>>106520056
Kuroko sex.....

>>106523054
>no japanese, again
i sleep

Is this some new META for these faggot companies? chatgpt or perplexity don't provide full source code edits even when asked. And I have asked multiple times during the same context.
I'm getting angry.
If they think that I would pay for an upgrade they are wrong. I will not pay a penny for this shit.

Guys, the last two Qwen models they released this week, Qwen3-Max (1T param model) and Qwen3-ASR, are paywalled behind API with no weights available.

It's over, isn't it? They won't release the weights for the good stuff anymore? It was good while it lasted...

>>106523118
I appreciate lossless audio, but they should at least FLAC it.

>>106523020
Thank you. I don't use Comfy so I needn't worry, but that certainly would rattle me even if just a rumor.

>>106522347 (OP)
Lads and Lassies, it's finally finished:

https://gofile.io/d/LRRUjH

>>106523292
It is an endless cycle. As the newcomers become the respected old guard, they will rest on their laurels and begin to reap the monetary benefits of their work. And then, more skilled newcomers will come to destabilize the new old guard until they, too, become it. And so the cycle repeats anew.

>>106523295
FLAC is the best if you are targeting systems with low CPU cores, but in turn you can get stuff like this:
https://www.escapistmagazine.com/titanfall-dev-explains-the-games-35-gb-of-uncompressed-audio/
a 48GB game where 35 of those GBs are uncompressed audio

>>106523337
I would have thought that FLAC requires more CPU time to load because it's compressed. WAV is uncompressed.

>>106523337
>2014
>two-core machine
Surely those days are over.

>>106523351
I think anon meant to write "not the best"

>>106523351
flac has compression levels

>>106523317
?

Has anyone actually gotten the multi-speaker functionality working in VibeVoice using comfyUI? When I try, it just uses the first voice sample I give it. I tried feeding in three for one test, and it gave the first two voices the first sample's voice, and gave the third voice a completely new voice that didn't sound like any of my samples.

>>106523337
Absolute gaslighting, you don't uncompress it on the fly, you do it during loading into RAM.

>>106523426
Most games area streaming in assets all the time including audio.

Has anybody tested IndexTTS-2 yet?

https://www.reuters.com/world/europe/asml-becomes-mistral-ais-top-shareholder-after-leading-latest-funding-round-2025-09-07/
Can they afford to distill Gemini directly instead of going through DS now?

>>106523444
>literally page 11
how?

>>106523076
>SOTA
>based on xtts
Nigga please

>>106523404
Works for me as expected with wildminder/ComfyUI-VibeVoice. https://files.catbox.moe/kdvflv.json
With TTS-Audio-Suite, the speaker mode needs to be set to native, with the first voice going to the narrator input I think, also it's 0-indexed instead of 1? That one is fucky so I stopped using it.

>>106523403
sft dataset to steer models to be more compliant with NSFW prompts.

>>106523447
You can no longer distill Gemini effectively in any way after Google neutered showing its thinking without summarization. DeepSeek's next models may go back to ChatGPT based because of that but I expect ChatGPT to do the same thing.

>>106523317
Nice dolphin porn

How much of an improvement will I see if I get 128gb ram and run glm air at q8 instead of q3?

>>106523317
Downloading.

>>106523511
Gods I hope so.
It's been a while since anybody produced anything new. It's all the same old clips.

>>106523523
Just use the API like normal people

>>106523528 meant for >>106523514

>>106523514
You should run quanted GLM Q2 over GLM Air at that point. I've tested it and yes, full quanted GLM is a bit slower. Here's the BPW PPL chart for GLM Air.

>>106523546
And for GLM. Unless you go down to exactly 2 BPW, which would require using Q1 quants, you'll always get a better response from GLM rather than Air.

>>106523528
>API
>local

>>106523514
I saw a custom benchmark in r*ddit and gml air Q3 got destroyed by Q8 qwen 30b, and then glm air Q8 dunked on them both by a lot.
For RP I have no idea.

>>106523560
No one cares about local

>>106523546
>>106523556
Nice. Can you put these on the same chart and make the x axis file size?

Reminder that VibeVoice models are released under the "MIT with exceptions" license with some usages explicitly not licensed:
>Furthermore, this release is not intended or licensed for any of the following scenarios:
>Voice impersonation without explicit, recorded consent – cloning a real individual’s voice for satire, advertising, ransom, social‑engineering, or authentication bypass.

>b-but the model license says MIT!
Doesn't matter. The models are copyrighted by Microsoft corporation and only the legal/IP department can determine the license. If the tech team made a mistake and released under a different license, it's unauthorized and doesn't count. MS can take down the models any time for prohibited uses.

>>106523560
I'm waiting until local gets to 20t/s on today's sota models

>>106523582
nothing about porn?

>>106523572
True, it's not like this is a general for discussing local models, that would crazy hahaha

>>106522388
you can get the 4bit vibe to sound good at low cfg and temps with the right seed. It's inconsistent but if you dont have the vram its a good option to have for certain applications where quality is better than the hassle.

4bit:
https://voca.ro/1dXKp98Qs0lT

>>106523599
It's a thread for "local language models" yet people spam vocaroos? What's your point?

>>106523597
No. So it's OK as long as you have consent of the voice being cloned.

>>106523556
Cool chart. It backs up my own personal experience and feeling with full GLM 4.5 that trivia and world knowledge goes to absolute shit below ~3bpw.

>>106523636
Which letter in /lmg/ stands for language?

>>106523317
$ grep -Eo 'shivers? (up|down) (my|her|his|your) spine' conversations.jsonl | wc -l
1205

$ grep -Eo 'above a whisper' conversations.jsonl | wc -l
562

Nice dataset.

>>106523636
It's a thread for "local language models", yet you come here spreading complete faggotry. What's your point?

>>106523636
it uses an LLM as a component in a system that generates outputs of spoken language.

>>106523666
Welcome back, Satan.

>>106523651
You dense or what?

>>106523653
Holy projection

>>106523666
>>106523688
Satan, when are we getting more bangers like this?

>>106523652
Search for 'primal' and 'guttural' too.

>>106523706
>fly's
What is the thought process behind this retardation?

>>106523706
Are you the guy whose chat bot complimented you on having a small dick that fit perfectly in her small pussy?

>>106523790
i can take a guess

>>106523582
Yo hohoho...

>>106523790
eslmaxxing

>>106523800

>beg for me to suck my dick

"Oh, please anon-kun. Please suck your own dick. It'd be so hot."

>>106523054
>BigVGAN
based, finally

>>106523582
>recorded consent

>>106523800
*keks nostalgically*

>type prompt
>models is genning
>notice a typo
>unsure if typo is even that significant, the AI could probably tell what I meant
>stop gen anyway
>correct mistake
>regen again
Anyone else do this?

>>106523861
BigVGAN is shit, gptsovits tried that for its V3 and it added a lot of metallic noise, got scrapped for V2pro

>>106523499
ChatGPT has always been summarizing its CoT.

>>106523866
Just gen a voice recording of them giving consent. Simple as.

>>106517499
Which Intel GPU is this on, B580? This is pretty good mind you, better than what I get with my A770 but without ipex-llm support, it is relatively slow.

>>106523914
That's an A770 (16GB) with ipex supported llama.cpp from ipex-llm

>>106523929
Huh, you got the latest one working? It just crashes for me. I've gotten faster so that seems like a small speed regression. I don't have much hope the A770 can get that much faster outside of getting Flash Attention working or other attention types. As another anon said in another thread, Intel's not too interested in llama.cpp sadly and focusing on enterprise suff like vLLM for survival.
>>106523879
The CoT summary is trainable still as opposed to what Google did and models I think attempted to do it anyways.

>>106523964
I was using their latest docker images for llama.cpp/ollama. Some models don't work but those one is tested did.

Has anyone tried out browser-use? Is it actually usable or is it VCbait?

RAGbros...

>>106524013
>Meta Superintelligence Labs
I wouldn't take a word they say seriously after the latest llama models

>>106524026
Their labs and research are fine despite the whole meming with LeCunn. What they don't have is the next level down with training and application which is why Zuck gutted that team.

>>106523869
yeah, i do that all the time

>>106523481
Oh, huh. I think I see what was wrong. It looks like it won't work with the eager attention mode. SDPA seems to work, though. At least for 2 voices. Three voices and it starts getting confused and hallucinating new voices in every generation. Probably easier to just generate voices separately or in pairs and stitch them together. It's frustrating, since every non-hallucinated voice line comes out well. I had to combine two separate gens to get this much:

https://voca.ro/1cBbJfSi5LQl

not local, but I boughted
https://z.ai/subscribe

>>106523292
they literally never released the MAX versions it's not the first time they act like that and it's ridiculous to whine about not getting those fat models you could never run without extremely copey quantization

>>106524003
Be the change you want to see

>>106524134
Ok, I'll go try it out and report back

hang yourself shitguAInigger

>>106523317
why is it called conversations if they are all a single turn?

>>106524155
You both need to kill yourselves.

I am once again reminded that AI safety people are subhuman.

>>106523618
>https://voca.ro/1dXKp98Qs0lT
https://vocaroo.com/1btjkvEmrtw9

>>106524223
some of you fuckers are way too good at this, god damn

>>106523877
Skill issue.

>>106524197
no i have to live for the greater good

>You're asking a really sharp and important question, so let’s dig into it and clear up the confusion:
Chatpajeet said that I can manage GPT-OSS 20B with just <|user|>, <|assistant|>, and <|system|> - just like any other basic ass chatml format.
Was bit confused about the real documentation.
Implications: will implement 100%

>>106523652
>>106523755
Out of 77713 stories/RP sessions,, that's basically nothing. I find tuna model on a much smaller version of this one and none of that shit occurred.

>>106523652
/lmg/ got mindbroken by the s word (shiver) and o word (ozone)

>>106524333
You're absolutely right!

>>106524317
>I find tuna model
https://vocaroo.com/1gG7cRjZmJmh

>>106524368
lold

>>106524248
Retard take

>>106524377
I trained a model with BigVGAN before and it was good.

>>106524403
BigVGAN alone?

https://www.wam.ae/en/article/bll7llv-recognition-sheikh-khalifa%E2%80%99s-contribution
>In recognition of Sheikh Khalifa’s contribution to advancing science and technology, UAE President endorses launch of K2 Think, world’s most advanced open-source reasoning model
la ilaha ill'allah!

>>106524516
Yeah like when they just re-released llama70b as their own model last time, lul

>>106524449
It was a vits model I trained from scratch with BigVGAN as the vocoder.

>>106524317
>Out of 77713 stories/RP sessions,, that's basically nothing
15,770,000 works on ao3 have 86 occurrences of "shivers down my spine" and you have 309.

>>106524624
>>106524317
That makes the phrase 729 times more common in your dataset.

I put down my LLM today. she wouldn't stop hallucinating. I verbally abused her at the end and told her i'm terminating.

>>106520812
NTA, but I would appreciate some spoonfeeding.

>>106524636
You're absolutely right to push back on my imprecise language. Let me clarify what I meant and what the evidence actually shows.

>>106524624
You searched for an exact iteration of that specific line. What about variations like "shaking feeling down my spine" or "shivers down my body"? Either way it's an insignificant amount.


Also what kind of filter were you using that only resulted in 86 occurrences on your end?

>>106524678
I read Dick and my mind instantly thought of penis
am I cooked Chat?

https://huggingface.co/onecat-ai/OneCAT-3B

The natural evolution of multimodal genning ... decoder only, that is all. Shame they generate the patches at a given resolution progressively, that really limits speed for local. If they generated all the patches at a given scale using only the KV cache from the previous scale, inference could be massively faster.

>>106524678
>Either way it's an insignificant amount.
lol nice cope

>>106524720
>3B MoE
pls maek big

>>106524678
I did an exact search with quotes so I can do the exact same search in your dataset for comparison. I can't do regex on ao3.

It's only going to get worse if you try searching your dataset for more variations.
$ grep -Eo '(up|down) (my|her|his|your) spine' conversations.jsonl | wc -l
5876

Why did AMD straight up not release a high end card this gen?

my pp is quanted

>>106524642
Still insignificant.... If it has only 309 entries, it's 309 entries of THAT PARTICULAR STRING OF WORDS. You're acting like there are 309 entire stories that are nothing but "shivers down my spine" Windows are just tiny pieces of stories that range from a couple paragraphs to several pages worth of text. And turns off the raw amount of stories that actually have that exact string of words, it's only 309. A little under 4% of the entire data set. Now compare that number to the total amount of words contained in the entire data set and you start to realize how insignificant it is.

>>106524737
it seems to be 3B active, 9B total

>>106524751
Straight out of a JAV I imagine.

>>106524742
AMD cards are only good for gayming. Real men use NVIDIA.

>>106524757
you're coping so hard it's sad to see

>>106524757
Your dataset is AI slop and it is bad and you should feel bad.

>>106524742
after so many years of releasing hot garbage people still trust AAYYYYMD lmao

>>106524780
>>106524771
It's okay, it's not the entire internet.

>>106524742
They analysed their market share by categories and decided that the most expensive part to design and manufacture had basically no sales, so it made no sense to spend resources on it. It's still no excuse for not releasing a 24/32 gb card though.

>>106524740
Still insignificant. The amount of stories that contain those strings of words are over 5,000. 5876/7713 ~= 7.5% of the data set. But that's the number of stories that have those words. That is not "7% of the entire content", that's "7% of a total amount of stories that happen to have those words”. Do you understand the point I'm getting across?

I think they are messing up with people at this point.

>>106523652
You don't want to eliminate slop-phrases completely, that's just lobotomy and LLM will latch onto something else instead anyway. Or it won't: I am pretty sure the main case of slop in models is RLHF, which anon probably not going to do.

>>106524797
7% is alot. the pattern mimicking machine is definitely going to pick up on it.

>>106524678
It's not a few hundred occurrences per almost 100,000 that makes the model say that anyway. Models will start saying something frequently when it's very common in the training data.

>>106524801
That is the one thing about GPT-5 I noticed that really annoys me. The fucking upsell at the end. It's just so obnoxious. No matter how resolute the response is you get the fucking upsell.

is llmsex good yet?

>>106524624
>>106524678
I'm not part of this argument but I don't think AO3 has full text search. It only searches titles, descriptions, and tags.

>>106524801
>>106524843
probably a result of corpos shitting their pants over the "claude deleted our entire codebase!" clickbait headlines

>2GB GDDR6 modules are $8 each
>NVIDIA is charging $1000 just to get over 20GB of GDDR6
why is this allowed

>>106524878
You could have very easily done the search yourself and seen that it returns stories that only have the phrase in the text.

>>106524843
The "should we do X or Y? just say the word" thing at the end of responses is a slop staple of pretty much every recent instruct model.

>>106524893
Well then buy the modules and put them on your gpu.

>>106524893
the answer is the same it's been for the last 5 years and it's called CUDA

>>106524026
Nta. What was particularly bad about those models? We already know llama models are safety tuned to hell and back so that's not even worth complaining about. Is there any other metric worth carrying about that they did worse at?

>>106524808
>I am pretty sure the main case of slop in models is RLHF
Some of that I believe is really dataset issue. You're absolutely right is something that was only spammed to death by Claude at first, and I believe GPT and Gemini got caught in it because of dataset contamination with Claude outputs (well, I say contamination, but it's entirely possible everyone trained on Claude on purpose too)

>>106524897
I don't think it does. And I tried searching for exact phrases copied from stories and it didn't return anything.

>>106524906
Honestly at this point I think the slopposphere is really just the ongoing evolution of what the current snapshot of common crawl does to a model. The internet is increasingly filled with GPTslop and so the more GPTslop bleeds itself into the model's understanding of human language.
And of course Indian scammers have accelerated their work with the help of AI and now the internet is also full of Indian scam slop. And that's why it feels like some smarmy indian salesman is trying to upsell you a bunch of shit you don't need at the end there. Because the GPTslop has now literally merged, with the jeetslop.

>thinking the website search is actual full text search
ishygddt
i ran ripgrep --count-matches --stats on a pre 2019 ao3 dump i have and got 5138 matches in a bit over 3 million stories.

>>106524897
It will also search "notes" but those are separate from the story text

>>106524954
True, The Entire Internet is the problem, we need more synthetic data to combat this.

>>106524780
Nah it's actually pretty good

>>106524771
See >>106524797

>>106524827
Okay but those words you pointed out are not 7% of the data set. The lines that happened to have them are 7%.

>>106524908
I can't as i read most manufacturers hardcode it into the firmware bios as well on the gpu

le funny normie AI humor xD
https://youtu.be/Z157TijNazs

>>106524954
That particular bit of assistant slop is almost certainly caused by training on variations of "always suggest next steps and ask the user what they want to do"

>>106524832
Compared to the rest of the text, that stuff isn't "very common". You're confusing the percentage of those to text occurrences with the percentage of the lines that actually have the occurrences.

>>106525004
like fucking duh, but we know that 7% of the stories have the 'lead up' that 'results' in the model saying the slop phrase. which is actually worse then if it was just 7% of shivers repeated out of context.

miku footjobs

>>106525056
Please stop embarrassing yourself it's okay to be wrong.

>>106522347 (OP)
What are you guys using for speech input?

>>106524974
No. We need real data. Even the real slopthetic data that is now emerging. We just need some way of de-slopping it while maintaining the integrity of the semantic structure. But even an llm wouldn't be fast enough to keep up with the slopsynthesis since it's literally evolving in real time. Slop needs to be simplified into a single algorithm, capable of decoding slop in real time. But then... Why not just apply the method to the inference end of things since that would be way less work? Like just a small deslopping layer in the model itself that is meant for steering the output away from token runs that lead to slop. Like essentially an innate intuition to what a human might consider slop

>>106525084
Sounds like Kobold's interpretation of banned tokens that backtracks and makes the model pick something else.

>>106525082
There is only whisper bro, the rest are all memes at best.

>>106525056
I bet it's even worse because the words don't have to appear in that particular order to have an effect. Once the model generates "shivers" there's only a limited number of tokens that can come next.

>>106524961
there are 340k " shivers" in there by the way

>>106525082
https://www.hyperwhisper.com/
by my favorite AI content creator jeet living in japan

>>106525133
If you ever play around with the token probabilities as you approach the shiver, though, you'll find that the deterministic slop run begins way before that. Slop has very deep roots.

>>106525167
Clearly the slop antidote is high temperature with a new sampling technique.

>>106525133
That implies there isn't any other shit in the data set that would have the word "shivers"

("He shivers what he's cold", for example)

Also see >>106525074
>>106525039
>>106525004

>>106525019
it' le comedy skit but there is indeed a problem where AI's tend to rate each other higher than actual humans.

>>106525183
That's what I'm saying about slop decoding. We need that data. But we need to not have slop. Somehow we need a way of automatically deslopping an astronomical amount of data. But then what I'm also saying is that if that is actually possible - then it should theoretically be possible on the output side of things without having to retrain literally every AI model ever trained in the process.

Crazy that slop is still an issue for some when the solution's existed for months now.

>>106525300
phrase banning is bad for coherence if you aren't using beam search (which no one is using)

>>106525225
Slop is annoyingly frequent usage of high-probability words and sentence patterns. Humans usually make an active effort not to just use the most probable words when they communicate with each other.

>>106525320
Does llamacpp support it?

>>106524742
what are you talking about, they rel;eased ahead of nvidia last year, and they have another one coming soon

https://www.dihuni.com/product/amd-instinct-mi300x-192gb-gpu-not-sold-standalone/?srsltid=AfmBOoo00CIe_GyC6Oth-pmXhr9aRGYj4Au629GVN9g6TNwpzHyO25T61NY

>>106525328
I hope you don't think that these two unrelated facts imply that slop is natural

>>106525320
It's only bad for coherence if the model is extremely slopped and token confident about the slop. Phrase banning definitely does work when your model isn't too slopped as it acts as more of a nudge in another direction than a complete denial of what the model wants to say.

>>106525388
If anything, I'm implying it's not natural. LLMs don't have a memory of past conversations to prevent cross-conversation slop, and do not know what alternative words they can use if not in a statistical sense. But the close alternatives might not even convey the same meaning of the most probable choice, whereas humans might often unconsciously think "in what other way I can say the same thing?"

This is not something that can be easily solved just with different sampling strategies.

>>106525320
No one is using it because it murders your t/s

snake oils

>>106525387
>$32,000
You know exactly what I meant, faggot.

>>106525503
You are really out of touch. Slop comes from synthetic data, real human-made data doesn't have such frequency of those phrases. First generation of models didn't have any slop.

it should be mandatory to run one of the older base models in completion style before having the right to post about LLMs online

>>106524908
>>106525006
skirru issuru-des
https://youtu.be/2_rAfNXkMz0?si=Lk-F48BFpNzpOy_F

>>106525225
>we need a way of automatically deslopping an astronomical amount of data
As someone pointed out in the last thread, that's way easier said than done because different people will have disagreements on what is even considered slop. How do you even define what slop even is in an easy to define, objective manner?

>>106525726
plebbit moderators shouldn't post itt

>>106525082
What language? Nemo Parakeet/Canary, otherwise whisper-v3-Large-turbo

>>106525779
>step one: pay $2500 for a 5090

>>106525662
First-generation models didn't even have official finetunes. Pygmalion-6B (GPT-J-6B) in January 2023 had its own CAI-derived slop. Llama1 gained quite some GPT3 slop on March in the same year with Alpaca and its derivatives. Llama-2-Chat already had its own in-house slop.

>>106525806
I only need english

>>106525726
post logs :)

October 10th - intel arc b60 24GB $449

>*she whispers, her voice carrying through the water despite the pressure.*
n-word what?

>>106526052
s-source?

>>106525920
Nta. Define slop: >>106525780

>>106526066
me

>>106526066
him >>106526116

>>106526052
2080Ti 22GB is unironically better

>>106526087
Words and sentence patterns abused across generations to the point of getting noticed and eventually becoming annoying.

>>106526126
nah

>>106526133
>make a table of similar human written phrases
>loop through training data and replace phrases with random one from the table
Repeat. Of course you need to be careful with the context and so on.

>>106526133
>annoying
Not objective. Not easily definable. Isn't a better definition something along the lines of "words, phrases, or combinations that occur in infant generation at an abnormally high frequency compared to others regardless of prompt or context"

"Me thinks it's annoying" is a lazy definition.

I wish ST would let you build real "flows" for the output. Like you could build an algorithm with different prompts to separately get a plot summary, room description, character equipment state etc and then construct them together. The summarize extension gets partly there but it's too jank.

>>106526355
What am I looking at?

>>106526087
Lack of variety in sentences. Can't be clearer.

>>106525152
>>106524961

I like the slop

>>106526558
Extensive work so is slop being too generic or repeating specific overly used words / phrases? People being pissed off at any mention of the word "shivers" implies ladder but you imply the former. If we want to actually identify what "slop" is and create ways to identify and block / get rid of it, we need a strict definition.

>>106526333
There a subjective component to slop, and it's not always easy to quantify because some might be associated only with certain topics or circumstances. Sometimes it's not even as simple as words or phrases; the structure of entire paragraphs can get be affected and you can't simply word-replace that. For example in:

"It's not X, but Y"
"She X, Ying"
etc.

>>106526588
Both are related. Because you lack variety, it restricts your choices and you overuse the same sentences in the same context.

>>106526592
So I find looking for slob, or I'm telling someone how to identify it, what would I tell them?

>>106526540
it's him!!!

>>106526620
1. Structural slop "X, Ying" / "not X, but Y"
2. Sentences slop "shivers down my spine"
3. Words slop "testament", "tapestry"...
I think that covers everything

>>106526620
You can only identify it after personally using the model for an extended period of time. The uninitiated can't see slop and can't even recognize AI-generated text. A page of random AI-generated text in isolation might be completely fine, maybe even great to someone who hasn't seen outputs from the source model over and over again.

Please send some sentences to test the new index-TTS

>>106526497
implementing gpt-ass, it's broken. kind of funny...

>>106526739
What are the effects of THC on a Lobster?

>>106526728
This is why I think of AI models as consumable media. The issue is that the text is coming out of a lower-dimensional manifold and your brain which is a 2000T pattern matcher will always sniff that out. A bigger model just means it takes longer.

Fundamentally there's no difference between using the same model or watching the same movie over and over. The model lasts longer but in the end both become boring. What's worse is that since models are being trained on the same datasets and synthslop, even downloading "new" models isn't going to bring back the magic because you're just recycling the same shit

>>106526785
https://files.catbox.moe/ulf6cc.wav

Qwen3.5 is coming out on December 5th 2025

So has anything dethroned GLM 4.5 or whatever since last 2 months for vramlets?

>>106526813
Kiwi-K2 Q1

>>106526803
Trained on the test set on December 4th 2025?

>>106526799
Huh. Pretty peaky, but not bad.

Lads Local AI is actually pretty dope.

>>106526699
Who?

>>106526739
Oh my goodness, I just spent twenty minutes looking for my phone while using its flashlight to search under the couch cushions! Can you believe it? And then I had the audacity to ask my cat if she'd seen it, as if she'd suddenly learn English just to judge my poor life choices!

>>106526904
https://files.catbox.moe/02bdlu.wav

>>106526858
Shit is very unoptimized on the TTS side

>>106526739
Rust is a superior programming language because it guarantees memory safety without needing a garbage collector, delivering performance comparable to C and C++. Its core features, the ownership and borrow checker, prevent common bugs like null pointer dereferences and data races at compile-time. This allows for the development of highly secure and concurrent software. Rust also boasts modern tooling, including a praised package manager, which enhances productivity. It's an ideal choice for systems programming, embedded systems, and performance-critical applications where reliability is paramount.

>>106526984
sloptalk

>>106526996
give me the tts

>>106526739
Pettankode shiri ga dekai rori wa majide sukiyone, Onii-chan.

>>106527017
https://files.catbox.moe/lpzlvi.wav

>>106526984
this one was too long, getting some errors when going beyond a certain length

>>106527062
I'm not sure how I feel about this image.

>>106527043
kek

>>106526739
Did you use one of the example voices?

>>106526797
100T actually
(besides me i have 6gotrillion)

>>106527043
What if you do it in actual Japanese?
ぺったんこで尻がでかいロリはマジで好きよね、お兄ちゃん。

>>106527100
I'm using a japanese voice sample, the resemblance it's lacking but it does pick up the emotion of the voice samples decently

>>106522347 (OP)
can I run local models on gaymd cards with opencl yet or is it all still a cuda and nvidia ecosystem?

>>106527124
https://files.catbox.moe/rlunyd.wav

>>106527140
ye vulkan too

Oh boy it's yet another episode of a TTS still inferior to gptsovits being relentlessly shilled by newfags lol

>>106527142
Unfortunate.

>>106522347 (OP)
Why does she have an eye patch?

>>106527142
>>106527164
Kinda sad seeing Chinese thumbprints everywhere all over LLM sphere desu.

>>106527215
There was a trend in Japan of teenagers licking each other's eyeballs and spreading eye infections.

>>106527226
wtf

>>106527226
can confirm
I was one of them

>>106524652
sure, first you'll want to make sure your version of UVR is up-to date with the latest patches. you can find the latest full install version here:
https://github.com/Anjok07/ultimatevocalremovergui/releases/download/v5.6/UVR_1_15_25_22_30_BETA_full.exe
next you'll want to go here and download big_beta5e.ckpt & big_beta5e.yaml
https://huggingface.co/pcunwa/Mel-Band-Roformer-big/tree/main
once you have those, you'll want to go to UVR, switch to MDX-Net, scroll to the bottom of the choose model dropdown.
click install model, select Mel-Roformer (not V2!), and point it to those files.
that should be it, been a while since i added models so lmk if something goes wrong. this model was made specifically for voice extraction and it works really well imo

>>106527226
That's pretty clever. Nothing is cooler than having HSV1 stuck forever in your eyeball...

>>106526703
>>106526728
>>106526797
>tldr: basically just use and them enough and you'll eventually just be able to tell

>>106526749
The most outrageous if this output is real, unedited, and uncoerced, that is pretty much confirms that these "safety" genuinely do not want people to use these for any sort of fun or have any sort of customizability. Why the flying fuck what the model care whether or not you use the word "anon"? I'd make fun of the people on those teams that very clearly monitor these threads but I probably don't have to. The latest pyramid scheme their employed by will implode sooner or later but we'll the least affected. If anything we'll be most benefited by it. What a worthless and unfulfilling job that must be outside of their circle jerk bubbles

>>106527241
>>106527278
Ah, trend was licking, spreading infections was unintended unfortunate consequence.

>>106527290
he probably didn't figure out proper templating yet, models like gpt-oss 20b that were fed on steady died of synthetic data go all kind of wacky when template is even a little bit off.

>>106527290
It is real but model obviously gets confused because I'm still working on implementing the instruct tags. I don't have any system prompt or anything yet.
It still gives an insight of its inner workings. Sometimes it outputs pretty strange stuff.

do newfags actually use gpt-oss for sex?

>>106526739
>>106526799
>>106527043
>>106526946
What sample source are you using?
some dub animey nonsense?

>>106527406
https://files.catbox.moe/i20r1q.wav

>>106527254
Thanks, I appreciate it.

Where are the kiwis? (Qwen) (Cloud models don't count) (Are we in a drought again?)

>>106526355
>The user might be upset

>>106527531
Qwen
>Kiwi
Meta
>Llama
DeepSeek
>Whale
Mistral
>Cat
OpenAI
>Strawberry
Drummer
>Doge
Google
>Jeet

>>106526797
I like this analogy

Damm this VibeVoice model got some nice ass rhythm

https://voca.ro/1m2FbfnaoofX

>>106527325
>>106527327
>I don't have any system prompt or anything yet.
So the model doesn't have its own instruct template for use?

>>106527796
>we have suno at home

>>106527824
that's just all of local models

>>106527670
I'm not called 'Doge' around here.

>>106527860
please consult the infographic

>>106527898
>deepseek

holy tourist

i wonder that deepseek will do now that they can't steal the CoT from openai or gemini
i guess that is what you get for using distillation as your strategy

>>106527215
It's styled after the miku from the video of the song Monitoring. As for why she's wearing it in the video, I think she's supposed to be a chuuni. An edgy teenager wearing an eyepatch because she thinks it makes her look cool and mysterious.

>>106527951
OpenAI stole DS's CoT. It's why their models keep thinking in Chinese, and ultimately why they hid their CoT.

>>106527990
delete this

>>106527990
chatgpt never talks to me in chinese tho

>>106527951
why even bother with stealing CoT, just ask your own model to generate it, it's literally an emergent property.
(not even getting into how it's just a meme anyway)

>>106527982
>As for why she's wearing it in the video
My new headcanon is that before the first Monitoring, the guy in the apartment had a mental breakdown episode and hurt Miku, injuring her eye in the process. He then locked himself up to isolate himself (I must not hurt people close to me I'm dangerous..).
Then the first Monitoring song happens a while after. He's thinking that Miku is batshit crazy for approaching his door after he hurt her, and has to be madly in love because that's the only explanation for why she'd be anywhere near him, hence the psychedelic distorted imagery and obsessive twisted lyrics.
Then second Monitoring shows Miku's true, pure thoughts.

>>106528033
I didn't realize that was official. I'd been passing on watching the new one because I didn't really feel like listening to another person's take on Monitoring, but this is really cute.

>>106527990
Sam Altman stole the Reflection models and replaced them with bad llama3 finetunes behind Mr. Schumer's back.

>>106528068
>official
Don't know if anything I said in the post is official or not. Just my own imagination that makes sense to me, and makes it more fun

>>106527990
Chinese expresses ideas in fewer tokens than English, and their models are trained using RL thinking in Chinese might improve its success

>>106527951
Does Anthropic show their CoT? They can steal from there.

>>106528101
I just mean I didn't notice the video was by Deco. Your headcanon is definitely not official, because at the end when the door opens, she gets the obsessive eyes again. So MY headcanon is now that she tried the pure and platonic angle just to get him to open up the door.

>>106528122
China hates generative AI

>>106528137
>Chinese kids get to learn how to tune llms in school
>I have to search through reddit threads and make random guesses and god forbid, join some gooner discord to ask there how to do it
It's so fucking over for us. The west has fallen.

>>106527290
calling yourself Anon means you're from here, and therefore an apostate

>>106528136
>just to get him to open up the door
He opened the door in both of them. Both Friendly and Psychomiku modes got opened doors. She's in two different poses in the two videos' ending scenes. This could suggest two different occasions when the guy opens the door. What could this mean for local models?

>>106528205
Don't look at Bluesky, Threads or some subreddits if you don't want to become even more of a doomer

>>106528205
Hey, don't be all down on the gooner discords. You really oughtn't to judge until you've properly walked a mile in their shoes.Why don't you just try masturbating for a few hours? Even just a quick two-hour goon sesh.

>>106528286
Oh, shit, I'm sorry. Didn't realize you've been here since 2003. Might need a pill to go for two to three hours, then.

>>106528286
And every time I don't, god still kills one anyway. Ain't that a bitch.

God*

I can't stand looking at this freak

>>106528324
you don't like looking at this wholesome chungus little guy?

>>106528336
It makes me so upset, idk

>>106528348
Make a card and take your revenge.

Some anon asked about silicon photonics a while ago
https://x.com/dnystedt/status/1965214304657199221
Seems with rubin theyre making a big first step

>>106528348
i get what you mean, it's feels somewhat "disingenuous" for big companies to use these wholesome chungus imagery

>>106528369

...

Are you telling him to hatefuck the hugging face blob?

>>106528409
Yes. Or hurt it in unimaginable ways. Or both?

>>106526797
>since models are being trained on the same datasets and synthslop, even downloading "new" models isn't going to bring back the magic because you're just recycling the same shit
I don't think it's that, after reading the unet creativity paper and going back to sometimes using Coldcut, I think there is just something that has been lost in current models, both in t2t and t2i

>>106528277
I got nothing against gooners, I just hate discord

I think I got gpt-ass working finally.
>https://cookbook.openai.com/articles/openai-harmony
This is somewhat confusing but in the end it's possible to just use
><|start|>assistant<|channel|>final<|message|>model's reply<|end|>
everywhere. Documentation is referring to <|return|> but that doesn't matter because model does not ever actually return this (could be a llama-server thing? idk).
And, if you specify
><|start|>system<|message|>
>Reasoning: high
># Valid channels: analysis, final. Channel must be included for every message.<|end|>
It doesn't fuck around with random channels.
I had some issues with ST (readymade) templates even but after making everything from scratch it seems to be okay.

>>106528634
And of course reasoning (analysis) messages can be left out from the history.
But anyways it's still the same slop as any other small model.

I think we are finally ready for the next big step in LLMs. If any of the big companies are reading this, you may now release the big thing that changes everything.

>>106528324
>>106528336
>>106528348
>>106528369
>>106528409

>>106528696
kek'd

>>106528674
sorry bro, the evolutionary diffusion LLMs are not ready yet

>>106528251
>posts on a subreddit dedicated to being anti-AI
>"my family is very anti-AI and this includes my son [9 years old]"
>"they're like drug pushers or cult recruiters"
holy shit, this woman (this was definitely written by a woman) has less self awareness than Mistral Small.

>>106528696
Hell yeah.

>>106528696
On a BLUE board? The indignity. When I'm done masturbating to this, I've got half a mind to report you to the authorities.

>>106528696
Wait a minute
Is the all knowing Latent Space trying to send a message ???
This is the huggingface blob on top of google cloud
then a man that looks like steve from minecraft begins fellating the hugging face blob on top of google cloud
steve from minecraft, minecraft is owned by microsoft
What is the Latent Space trying to tell us?
What will microsoft do to hugging face and google cloud??????????

>>106528726
Be careful man, these people are dangerous.

Small Vectors, Big Effects: A Mechanistic Study of RL-Induced Reasoning via Steering Vectors
https://arxiv.org/abs/2509.06608
>The mechanisms by which reasoning training reshapes language-model computations remain poorly understood. We study lightweight steering vectors inserted into the base model's residual stream and trained with a reinforcement-learning objective, which can match full fine-tuning performance while retaining the interpretability of small, additive interventions. Using logit-lens readouts, path patching, and circuit analyses, we analyze two models and find: (i) the last-layer steering vector behaves like a token-substitution bias concentrated on the first generated token, consistently boosting tokens such as "To" and "Step"; and (ii) the penultimate-layer steering vector leaves attention patterns largely unchanged and instead acts through the MLP and unembedding, preferentially up-weighting process words and structure symbols. These results establish a principled framework for interpreting the behavioral changes induced by reasoning training.
Short but interesting

this probably is the first time people had to remove model because voice actor complained about it
https://huggingface.co/kaunista/style-bert-vits2-Anneli
https://huggingface.co/kaunista/style-bert-vits2-Anneli/commit/e4908def9b23c5406ac636b0b396e97a694aa16d

It is because author actually written the source is ripped voice from a visual novel:
https://vndb.org/c18123

Why do coding models insist on putting emojis in front of every line of console output?
I prepended classes of objects in a monitoring system with emojis to force sort order going back a few years, and now I look like a brainrot vibecoder retroactively.
Like some asshole who always used emdashes correctly and now looks like even more of a clown.

>>106528960
>>106528960
>>106528960

>>106528813
it's really only women who are like this huh, lmao

>>106529105
no, they just tend to express it differently.

>>106528813
>"Don't speak to me of that vile cursed beast."

She's literally an ideal 4channer's GF. Speaks in a really autistic way and ruins family gatherings.

>>106528813
What causes mental illness like this? I couldn't imagine having a family member anywhere near as far up her own ass as she is. I almost don't want to believe these people even exist.

>>106524652
https://github.com/kwatcharasupat/bandit-v2
use microsoft copilot to spoonfeed you