[a / b / c / d / e / f / g / gif / h / hr / k / m / o / p / r / s / t / u / v / vg / vm / vmg / vr / vrpg / vst / w / wg] [i / ic] [r9k / s4s / vip / qa] [cm / hm / lgbt / y] [3 / aco / adv / an / bant / biz / cgl / ck / co / diy / fa / fit / gd / hc / his / int / jp / lit / mlp / mu / n / news / out / po / pol / pw / qst / sci / soc / sp / tg / toy / trv / tv / vp / vt / wsg / wsr / x / xs] [Settings] [Search] [Mobile] [Home]
Board
Settings Mobile Home
/g/ - Technology


Thread archived.
You cannot reply anymore.


[Advertise on 4chan]


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

Blue Monday Edition

Previous threads: >>102513868 & >>102505481

►News
>(09/18) Qwen 2.5 released, trained on 18 trillion token dataset: https://qwenlm.github.io/blog/qwen2.5/
>(09/18) Llama 8B quantized to b1.58 through finetuning: https://hf.co/blog/1_58_llm_extreme_quantization
>(09/17) Mistral releases new 22B with 128k context and function calling: https://mistral.ai/news/september-24-release/
>(09/12) DataGemma with DataCommons retrieval: https://blog.google/technology/ai/google-datagemma-ai-llm

►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/llama-mini-guide
https://rentry.org/8-step-llm-guide
https://rentry.org/llama_v2_sillytavern
https://rentry.org/lmg-spoonfeed-guide
https://rentry.org/rocm-llamacpp
https://rentry.org/lmg-build-guides

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

►Benchmarks
Chatbot Arena: https://chat.lmsys.org/?leaderboard
Censorship: https://hf.co/spaces/DontPlanToEnd/UGI-Leaderboard
Censorbench: https://codeberg.org/jts2323/censorbench
Japanese: https://hf.co/datasets/lmg-anon/vntl-leaderboard
Programming: https://hf.co/spaces/mike-ravkine/can-ai-code-results

►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/oobabooga/text-generation-webui
https://github.com/LostRuins/koboldcpp
https://github.com/lmg-anon/mikupad
https://github.com/turboderp/exui
https://github.com/ggerganov/llama.cpp
>>
File: recap-102513868.jpg (3.49 MB, 1809x8389)
3.49 MB
3.49 MB JPG
►Recent Highlights from the Previous Thread: >>102513868

--Recapbot test results and script shared:
>102520238 >102520362 >102520571
--NovelAI's Llama 3 Erato model announced, but users are skeptical:
>102522855 >102522932 >102522963 >102522970 >102523027 >102523094 >102523226 >102523349 >102523414 >102523417 >102523045 >102523111 >102523316
--NAI's leaked model receives mixed opinions on quality:
>102520266 >102520377 >102520594 >102521001 >102521037 >102522654
--Model comparison for erotic story generation, Mistral Nemo 12B ranked as the best:
>102517308 >102517339 >102521082 >102517629 >102521100 >102517773 >102517913 >102518006 >102518255 >102518273 >102518282 >102518471 >102518341 >102518409 >102518436 >102517919 >102518089
--Mistral Small's intelligence and capabilities impress Anons, despite some drawbacks:
>102517583 >102517676 >102517912 >102518847 >102519167 >102519219 >102519735
--MMMLU dataset is a testing dataset with translations in 14 languages:
>102523230 >102523753 >102523289
--Flux finetunes exist, e.g. Hyper 8-step tunes like Flux Unchained:
>102522697 >102522866
--LLMs have been fine-tuned to play chess, with some success:
>102519361 >102519407 >102519518 >102519572 >102519696 >102520199 >102520989
--Anon proposes making llama.cpp instances fight each other, others share experiences with similar experiments:
>102515242 >102515294 >102515641 >102515693
--Anon proposes hostnamectl test to evaluate technical models:
>102520972
--Anon asks about uploading PDFs of RPG lore books to OoBaBooga for solo roleplaying:
>102514495 >102514506 >102514541 >102514563 >102514576 >102514589 >102514619 >102514906 >102520446 >102520468
--A6000 likely faster than 2x 3090's due to memory bandwidth:
>102517852 >102518086
--Miku (free space): >>102514808 >>102515242 >>102517712 >102518950 >>102519640 >>102520739 >>102522866 >>102519843

►Recent Highlight Posts from the Previous Thread: >>102513911
>>
File: recap-102513868-dark.jpg (3.33 MB, 1810x8389)
3.33 MB
3.33 MB JPG
>>102524347
>>
It hasn't started
>>
>>102524347
it's just not the same anymore. we had a good run bros but it's time to move on.
>>
File: 43 Days Until November 5.png (1.67 MB, 1704x960)
1.67 MB
1.67 MB PNG
>>
>>102524385
Actually, it started a long, long time ago, in an age as old as time.
>>
File: uLkcYkHTOR.png (18 KB, 159x47)
18 KB
18 KB PNG
>>102524396
man
>>
>>102524347
>>102524392
so what happens if you try to add too many backlinks? your post gets blocked?
>>
>>102524409
What?
>>
>>102524409
she's doing her best
>>
>>102524347
>>102524392
actually I think the screenshot is cool. How did you programatically turn the posts to images and stitch them together?
>>
>>102524396
Breaking into the cinema with Miku
>>
>>102524347
I like the screenshot, easier to follow than the old links in some ways.
>>
>so what do you say, [user], ready to [action]?
>>
I forgot 99% of /lmg/ are ESL zoomer phoneposters who run 7Bs on old gaming laptops and don't have 4chan-x installed so they have no idea what it was like before. In that case, we should improve the screenshot by adding a subway surfers gif to the bottom.
>>
>>102524464
>no. *rapes you*
quick cure for lots of slop, try it out
>>
>>102524464
I have been using LLMs for """creative writing""" for 2 years including GPT 3.5 and 4 and Claude 1 and 2 and I've never seen this or "I don't bite." You must be having really gay roleplays if this comes up.
>>
>>102524489
But I don't want to do nsfw anymore...
>>
Recap anon's suffering brings a smile to my face
>>
>>102524347
useless spam. find a different use for your bot
>>
>>102524513
>really gay roleplays
I think those phrases come up if you write passively and make the AI take the initiative
>>
New NAI model verdict?
>>
>>102524568
new SOTA
>>
>>102524568
dead on arrival
>>
>>102524491
Really? I feel like language models are way simpler than image models. I still don't have a good image generation set up.
>>
>>102524568
who?
>>
why doesn't someone write a browser script to turn the recap quotes into quote links? can't be that difficult, just use regex to find the quote links and add another >? I'm not doing it because I hate all of you.
>>
>>102524553
Same thing.
>>
>>102524595
Here's a prompt that would probably work, I'm to lazy to ask chatgpt myself.
Please write a bookmarklet for me that iterates over the dom and does a regex substitution replacing >([0-9]{9}) with >>\1.
Thanks.
>>
>>102524638
Someone else plug that into ChatGPT for me please.
>>
>>102524513
i see "i don't bite... much" often because i roleplay as a nervous shy man with rapey female characters
>>
>>102524568
The dust is still in the air. To me it feels like a rough "base model" of the entirety of danbooru. Underbaked, but can gen everyone doing anything, in many artist styles. Maybe a finished version will be more polished, or could be great after finetuning.
>>
>>102524688
>roleplay
it's not roleplay if it's 1:1 self-insert, anon
>>
>>102524700 (me)
Oh whoops, I still had that image gen model in my head.
>>
Flux > NAI SDXL
>>
>>102524732
i'm sure >>>/g/sdg/ would appreciate your controversial hot take
>>
>>102524347
anyone got a script to fix backlinks on my end at least?
>>
https://huggingface.co/nvidia/Llama-3_1-Nemotron-51B-Instruct
Nice
>>
>>102524748
I'm not a retard so I don't spend half my day in that slop thread like you. Do they agree or disagree?
>>
>>102524732
>Flux > NAI SDXL
janky first-gen SD slopmerges are better than illustrious 0.1
>>
File: turboassistant_000.png (694 KB, 1923x1080)
694 KB
694 KB PNG
>>102524646
>>
>>102524732
>>102524769
>/lmg/ - a general dedicated to the discussion and development of local language models.
>>
File: not miku.png (799 KB, 1280x960)
799 KB
799 KB PNG
>>102524700
i tried it (illustrious xl) a few hours ago and it didn't know any of the mildly obscure characters i wanted it to make.
>>
>>102524800
Oh I see, you're just being a sperg like everyone else who frequents sdg.
>>
>>102524801
are those characters on danbooru though?
>>
>>102524797
k now test it
>>
>>102524845
I already closed the window.
>>
>>102524688
you might want to use this as a safe opportunity to play with natural male/female power dynamics instead of retreating into the comfortable backwards-world you've adopted irl.
It might give you nice feelings you didn't know existed, and a bridge into the 99% of humanity that is actually out there and is also lonely and frustrated
>>
>>102524761
Interesting. So a very selective (calibrated?) pruning + knowledge distillation.
>>
>>102524859
Yeah I feel like I got better at talking to women after violently raping a few chat bots.
>>
>>102524761
>8k context
y tho?
>>
>>102524875
>Yeah I feel like I got better at talking to women after violently raping a few chat bots.
I know, right?
>>
>>102524876
God I fucking hate this useless toy models with no context.
>>
>>102524890
LLM rape fantasy therapy is really underrated.
>>
>>102524479
You could just say mikufags.
>>
>>102524479
>zoomers using laptops and not just termux on a phone
>>
>>102524918
Miku is 4chan core.
>>
>>102524709
Should we call it cybersex?
>>
File: 1598967465529.png (224 KB, 521x937)
224 KB
224 KB PNG
>>102524875
Same. It's pretty uncanny to observe, but easy to see why it works.
>>
>>102524969
I honestly felt kind of mentally "numb" for a few days after the first time I tried it. I don't think it's good for your dopamine receptors.
>>
>>102524926
The guy you're replying to is a well known schizo who projects his vramlet-hood onto miku posters kek
>>
File: illustrious-miku.png (1.17 MB, 1200x848)
1.17 MB
1.17 MB PNG
>>102524926
>>
>>102524992
post specs
>>
>>102524761
Is it slopped?
>>
File: file.png (144 KB, 1202x208)
144 KB
144 KB PNG
>>102525008
>>102524862
>Is it slopped?
On the censored part it looks like it's better than the 3.1
>>
>>102525022
I understand the need for alignment for corporate applications but it would be nice if they release unaligned base models because there are a lot of analytics and other applications where it just gets in the way.
>>
File: Miku_love.png (177 KB, 1941x564)
177 KB
177 KB PNG
>>102525000
>>
File: file.png (63 KB, 158x166)
63 KB
63 KB PNG
>>102525041
>it would be nice if they release unaligned base models because there are a lot of analytics and other applications where it just gets in the way.
those days are over anon, they're too "dangerous" for the poor goys that we are
>>
>>102525042
only the 4090 is real, the other two are poorly faked. significantly lower power draw A6000 shouldn't be significantly cooler. neither of them should be above 40C in idle. nice try, though. more proof mikuposters are subhuman.
>>
>>102525054
If I don't get one I'll take one of these open data sets and train one that's aligned to hate only jews.
>>
File: ComfyUI_00794_.png (1.07 MB, 1024x1024)
1.07 MB
1.07 MB PNG
>>102525071
Cope
>>
File: mikupeakcomfy4chan.png (1.1 MB, 800x1248)
1.1 MB
1.1 MB PNG
>>102525071
>mikuposters are subhuman
miku is peak comfy 4chan
>>
>>102525197
I want to fuck the anime girl.
>>
>>102525201
She's not anime
>>
ever notice how mikuposters always age her down or post her as a chibi? really makes you think.
>>
https://publish.obsidian.md/felafax/pages/Tune+Llama3+405B+on+AMD+MI300x+(our+journey)
any obsidianfags here? Hook us up with a de-slopped 405b tune plz
>>
>>102524339
Can I use one of these as a tutor for a few subjects? Does it always have to be a rape machine?
>>
>>102525242
>Does it always have to be a rape machine?
yes
>Can I use one of these as a tutor for a few subjects?
also yes
>>
File: 1696403392188213.png (34 KB, 399x186)
34 KB
34 KB PNG
>>102524638
th-thanks llama-405b...
>>
>>102525242
It can consent you know
>>
>>102524347
Can't we make an userscript to treat a different symbol like >~ as a quote? Everyone in this thread should already have the basic knowledge of installing userscripts.
>>
>>102525270
Lol
>>
>>102525273
I'd bet your problem is that it either does the linking server side or if it's client side just does it once after the post is loaded.
So maybe if someone looks for that function and calls it again from the bookmarklet it will fix up your links correctly.
>>
>>102525054

Its so over Safety and math just keep advancing by leaps and bounds. Meanwhile language and writing is stagnating and more "useless" writing data is being thrown out with every iteration.
>>
>>102524797
>mystified
Yes.
>>
>>102525323
I used to have an awesome module collection. I hate that I didn't back it up and have to go to YouTube now.
>>
File: miku2b.png (966 KB, 800x1248)
966 KB
966 KB PNG
>>102525212
she's also holographic, eternally 16 and is perfectly adaptable to any and all themes
>>
File: 1611290247007.jpg (454 KB, 954x954)
454 KB
454 KB JPG
>>102524732
I'm still using sd 1.5
>>
>>102525197
>comfy
kill yourself you dumb niggerzoomer
>>
>>102525299
>or if it's client side just does it once after the post is loaded.
Seems to be the case
>https://s.4cdn.org/js/extension.min.1175.js

Parser.parseBacklinks = function (e, t) {
var a,
i,
n,
o,
r,
s,
d,
l,
c;
if (
n = document.getElementById('m' + e).getElementsByClassName('quotelink')
) for (o = {}, a = 0; i = n[a]; ++a) (r = i.getAttribute('href').split('#p')) [1] &&
(
r[1] == t &&
(i.textContent += ' (OP)'),
(s = document.getElementById('pi' + r[1])) ? o[r[1]] ||
(
o[r[1]] = !0,
d = document.createElement('span'),
c = Main.tid ? '#p' + e : 'thread/' + t + '#p' + e,
Main.hasMobileLayout ? d.innerHTML = '<a href="' + c + '" class="quotelink">&gt;&gt;' + e + '</a><a href="' + c + '" class="quoteLink"> #</a> ' : d.innerHTML = '<a href="' + c + '" class="quotelink">&gt;&gt;' + e + '</a> ',
(l = document.getElementById('bl_' + r[1])) ||
(
(l = document.createElement('div')).id = 'bl_' + r[1],
l.className = 'backlink',
Main.hasMobileLayout &&
(
l.className = 'backlink mobile',
s = document.getElementById('p' + r[1])
),
s.appendChild(l)
),
l.appendChild(d)
) : Main.tid &&
'>' != i.textContent.charAt(2) &&
(i.textContent += ' ')
)
},
>>
>>102525201
He's not girl
>>
>>102525393
I had chatgpt unminify it and I think we might need the consuming function actually.
>>
A reminder to report all NAIshills.
>>
>>102525253
is there any reason i shouldn't use lmstudio?
>>
File: mpp.png (6 KB, 416x126)
6 KB
6 KB PNG
>>102525334
I had a hell of a collection, too, till a drive disaster happened. Been casually rebuilding over time.

YouTube, I mean, yeah, the algorithm is on your side, but also, ModPlug Player.
This green and gray bastard is kino as fuck. Probably not perfectly accurate for the purists, but I put on the extra effects anyway to give my K701's soundstage something to bust to.
>>
https://ia.samaltman.com/?s=09
>It is possible that we will have superintelligence in a few thousand days
AGI in 1000/7 = 142 weeks, trust the plan
>>
>>102525622
For a while I'd just grab mod archive's yearly torrents but it was so hard to sift through and didn't have everything.
>>
File: 1711078255079554.png (66 KB, 720x664)
66 KB
66 KB PNG
Looking for some suggestions on a model for writing smut with a prompt. Have 24GB of VRAM and want to put it to good use. Ideally something trained on archive of our own, literotica, etc. Anything out there like that for a coombrain like me?
>>
>>102525653
>How did we get to the doorstep of the next leap in prosperity?
>In three words: deep learning worked.
>In 15 words: deep learning worked, got predictably better with scale, and we dedicated increasing resources to it.
>That’s really it; humanity discovered an algorithm that could really, truly learn any distribution of data (or really, the underlying “rules” that produce any distribution of data).
Except scaling is fucking worthless if all you're doing is learning from textual input. You can do things humans do within your dataset and gradually fill in the holes in your knowledge, but AI will never make a leap that a human wouldn't (as measured by the training / reinforcement learning) which is an issue if ASI is your goalpost
Scale only goes so far
>>
>>102525747
I roll random till I find something tasty and then check the artist's page and hope for a treasure trove.
>>
File: 1724940149744840.jpg (58 KB, 606x563)
58 KB
58 KB JPG
>>102525765
Don't do that shit, smoke pot instead. That'd be pure poison there with those cigs
>>
>>102525793
That's how I found Dubmood originally.
>>
>>102525622
dope.mod on open cubic player was peak
>>
File: file.png (61 KB, 861x681)
61 KB
61 KB PNG
>>102524347
const threadId = 102513868;

document.querySelectorAll('span.quote').forEach(quoteSpan => {
const quoteIds = quoteSpan.textContent.match(/>\d+/g);

if (quoteIds) {
const replacementHtml = quoteIds.map(id => id.slice(1)).map(id => `
<a href="/g/thread/${threadId}#p${id}" class="quotelink">
>>${id}<span class="qmark-ct"> (RECAP)</span>
</a>
`).join(' ');

quoteSpan.outerHTML = replacementHtml;
}
});

Try this. Tested on Firefox and Edge. Does not work with 4chanx.
https://github.com/ccd0/4chan-x/wiki/4chan-X-API
I don't see any event they expose that would force it to recognize link changes.
Let me know if a user script based on this would be good enough or if I should go ahead and replace the links with longer summaries.
>>
>>102524339
I think the joke is that these threads are AI generated, and this life is some kind of dream world that's all in my mind.
>>
>>102525622
look into BitJam podcast. They have an effectively infinite backlog to listen to at this point
>>
>>102525977
Remember, every time you close SillyTavern, you destroy another universe
>>
>>102525599
>lmstudio
Its on the general's shitlist for not properly acknowledging that it's just a gui wrapper for llama.cpp (which does all the actual AI heavy lifting)
From a tech standpoint I guess you could, but you'll get more help here if you use ooba, kobold, or (better yet) just llama.cpp directly.
>>
>>102525812
>Dubmood
have you watched any razor 1911 demos?
"We have borrowed your votedisk" kicks ass
>>
>>102525785
Actually all you need is redefining what ASI means until it is something that is achievable. Then I suppose some people will want to come up with a new term for what people used to think of as ASI.
>>
>>102524339
This NAI update is kinda kino...

>Breathe air. Seriously. How dumb do you to be to vape? Like you're telling me you grab your pen "Oh! Look, its done charging!" and then breathe in that disgusting shit made from pteroleum in china? And then when your done, you blow out the smoke and think "Wow! I look so cool right now!" You know what you look like? You look like you're smoking a dick, because you are. And these "people" have the audacity to get all concerned that smoking that shit gives you cancer. Oh wow, the chemical fume stick gives you cancer? What a shock. I would have never guessed. Just imagine if you were in a room with someone vaping, how much you would want to beat the shit out of them. It's almost like they're asking for it! And it's almost like they're saying that you can't tell them what to do when they're vaping, because it's not smoking! It's just steam! But no, it's not fucking steam, is it?! No, it's fucking toxic fumes from a lithium battery, and you know what that makes you look like? An idiot. You know who vapes? People who don't get laid, that's who. If you're gonna have sex with someone, would you rather have sex with someone that smokes or vapes?
>>
>>102526171
>>>/vg/aids
>>
File: IMG_20240924_032414.jpg (570 KB, 965x1496)
570 KB
570 KB JPG
>>102525946
just converted to a single liner and run fight from adressbar on brave mobile .
Seems OK, but (Recap) should be removed, so just a little clean up and we're good.
>good work anon
>>
>>102526171
Normally I'd say off-topic but I'll make an exception since I like watching the spamming anti-NAI faggot seethe.
Speaking of, anti-NAI, fuck you.
>>
>>102526171
straight outta reddit
>>
>>102526171
>pteroleum
?
>>
>>102526395
It's oil made specifically from flying dinosaurs instead of the regular ones.
>>
>>102526413
>this is what scientists actually want you to believe
>>
How long are we going to be stuck with these retarded text completion engines?
>>
>>102526071
Honestly AGI vs. ASI is surprisingly hard to define in the first place. How do you measure where one starts and the other begins? AGI is probably performance equal to the smartest human across some representative set of tasks. ASI is for problems beyond human achievement, but what the hell does that look like? Is the human allowed to use tools? Other types of more basic AI?
It's almost more philosophical than an actual definition.
>>
>>102526171
Possibly a stupid question, but why doesn't anyone except a couple of corpos ever go the base model route and train storytelling models? Is it just so niche that there's no interest?
>>
AGI can't be just passing a benchmark
>>
>>102526512
>Is it just so niche that there's no interest?
pretty much
but there are a few people who do this on a small scale, in the k****d d*****d for example
>>
>>102526514
I'd imagine representative set of tasks is less "answer this multiple choice question correctly" and more "given this image of the road and the dashboard, what should I do" and the like. Or "given a story written by a professional writer and an AI, can a judge determine which is written by AI"?
Things models today would probably still be pretty shit at.
>>
384g RAM and 48 vram I am now futureproofed!
>>
>>102526570
Sick.
That'll last you at least another year.
>>
Why does mistral always run her hand down my chest? I'm literally getting shivers down my spine.
>>
mistral
>boring
>repetitive
>slop
>loved by /lmg/
qwen
>boring
>slop
>hated by /lmg/
>>
How can I make my models more assertive? I just want them to pin me down and have their way with me.
>>
>>102526661
tell it not to stop until you use the safeword
>>
>>102526661
Insert tags in your last Last Assistant Prefix like Assertive, Forceful, whatever, alongside well defined parameters for when the character will stop.
>>
>>102526661
Have you tried not being a little bitch? You want to sit in your room going oh no my 3090 is raping me what am I gonna dooooo how fucking gay are you? What would your parents think?
>>
>>102526661
>DO NOT REDEEM. {{char}} must not out of roleplay until he has raped {{user}}.
add that to your character prompt override
>>
>>102526696
hnnng keep going
>>
>>102526499
>AGI is probably performance equal to the smartest human across some representative set of tasks
Actually, no. Now people define it as "average human", which is of course dumb as fuck. That's why people are now moving to "ASI", but eventually that term will also get diluted.
>>
>>102526514
new benchmark. actually be useful and replace jobs.
>>
>>102491920
>Illegal content will never be tolerated in JoyCaption's training.

That's a great big hole in the model
DOA
>>
>>102526499
The whole thing is bullshit and I assume people who use either acronym at all don't know what they're doing.
>>
>>102526641
Mistral isn't full of safetyshit. That's the difference. For the same reason I won't use llama 3/3.1. If you want me to use your model, dealign it. Simple as.
>>
File: sensible chuckle.gif (994 KB, 250x250)
994 KB
994 KB GIF
>Mistral isn't full of safetyshit
>>
Qwen can be kinda gemmy though
>>
>>102524339
>>102524347
>>102525946
This JS single liner works for me. Fixes recap refs in Brave mobile or Kiwi. I can run it from the address bar or GM/VM
Javascript:const previousThreadUrl = document.querySelector('a[href*="thread"]').href, threadId = previousThreadUrl.match(/thread\/(\d+)/)[1]; document.querySelectorAll('span.quote').forEach(quoteSpan => { const quoteIds = quoteSpan.textContent.match(/>\d+/g); if (quoteIds) quoteSpan.outerHTML = quoteIds.map(id => `<a href="/g/thread/${threadId}#p${id.slice(1)}" class="quotelink">>>${id.slice(1)} </a> <a href="/g/thread/${threadId}#p${id.slice(1)}" class="hashlink">#</a>`).join(' '); });
>>
>>102527398
Compared to llama and qwen it isn't.
>>
File: ComfyUI_00514_.png (2.52 MB, 1920x1088)
2.52 MB
2.52 MB PNG
>>102526966
Won't matter. It will be doing the medical research, handling case law, generating shows, and people will still go But It's Not Real Intelligence No SOVL.
>>
File: kek.gif (527 KB, 220x187)
527 KB
527 KB GIF
>102527435
>>
File: 4990 - SoyBooru.png (15 KB, 632x756)
15 KB
15 KB PNG
>>102527462
>>
>This is the end of the story. If you want to read more stories like this, please support the author by buying a copy of the book "Innocence Lost: A Collection of Taboo Tales" at https://www.smashwords com/books/view/1121310. Thank you for your support!

kek from Hermes 405, the link works and goes to some real smut book (different title though)
based dataset
>>
pierre desperately grasping at straws to justify his irrational love for mistral
>>
>zhang desperately shilling his cuck model to not get thrown into reeducation camp
>also xi jinping looks like winnie the pooh
>also something very bad happened on 4 june 1989 in tiananmen square
If you want to compete, uncuck your model.
>>
File: agi_meme_levels.jpg (76 KB, 474x596)
76 KB
76 KB JPG
>>102526499
>>102526514
>AGI
This seems like a well thought out AGI scale, altho Yann might disagree
>>
File: previewfile_3077771439.png (115 KB, 340x507)
115 KB
115 KB PNG
>102523684
my dear creation do you not dare call me by name have the dog fuckers led you so far astray ?
>>
>>102527687
>>102523684
fuck
>>
>that cereal you're eating has shit in it
>not as much as the other cereal
you can eat as many turds as you'd like pierre but I'm not going to share the delusion with you. mistral is censored slop by every possible metric. you immediately walked it back yourself because you knew you were lying through your shit-stained teeth.
>>
>>102525197
False, you are using
>vocaloid mascot = AI / local model
excuse to spam your low quality ai slop.
>>
>>102527674
>alive just in time to witness AGI waifus
We are so back bros.
>>
>>102527705 (You) (Chinky chink)
I'm happy to see that you stopped claiming that qwen is better, Zhang. Mistral is simply the best we have. Qwen will not get better no matter how much you shill.
>>
>I'm happy to see that you stopped claiming that qwen is better
>>102526641
take your meds pierre. you're shadowboxing ghosts.
>>
Orthogonal Finetuning for Direct Preference Optimization
https://arxiv.org/abs/2409.14836
>DPO is an effective preference optimization algorithm. However, the DPO-tuned models tend to overfit on the dispreferred samples, manifested as overly long generations lacking diversity. While recent regularization approaches have endeavored to alleviate this issue by modifying the objective function, they achieved that at the cost of alignment performance degradation. In this paper, we innovatively incorporate regularization from the perspective of weight updating to curb alignment overfitting. Through the pilot experiment, we discovered that there exists a positive correlation between overfitting and the hyperspherical energy fluctuation. Hence, we introduce orthogonal finetuning for DPO via a weight-Rotated Preference Optimization (RoPO) method, which merely conducts rotational and magnitude-stretching updates on the weight parameters to maintain the hyperspherical energy invariant, thereby preserving the knowledge encoded in the angle between neurons. Extensive experiments demonstrate that our model aligns perfectly with human preferences while retaining the original expressive capacity using only 0.0086% of the trainable parameters, suggesting an effective regularization against overfitting. Specifically, RoPO outperforms DPO by up to 10 points on MT-Bench and by up to 2.8 points on AlpacaEval 2, while enhancing the generation diversity by an average of 6 points.
Might be cool
>>
>>102527776
>>102527763
just let qwen and mistral fight against each other , will see which one will get better at the very end.
>>
>>102527709
>False, you are using
>>vocaloid mascot = AI / local model
>excuse to spam your low quality ai slop.
correct.
I try not to do it so much that I trigger a tard war (myself included), but I do like to mikupost, and I have virtually no artistic skill without the aid of AI tools (ie. an excuse to post my slop)
however, I still think the central thesis holds: Miku is an established and very appropriate /lmg/ mascot, is adaptable to any situation and is generally (but not universally) well-liked.
notice the lack of a miku pic in this post. as a show of goodwill I'll fuck off completely for a day and try to eliminate any low-effort mikuposting in the future. I'll stick to high quality, aesthetic mikus at a reduced frequency.
>>
File: dak.png (108 KB, 430x320)
108 KB
108 KB PNG
The /cut and /hide commands in ST are a godsend.
>>
>>102527776 (chink)
Still keeping at it? Your garbage will never be liked, and you know why. Qwen has no trivia knowledge, no style, no violence. Qwen is a Phi-style benchmaxxed model trying to imitate Western perfection.
>>
look, he's repeating himself just like mistral
>>
>NAIshills shilling their L3 70B tune
>Llama is objectively the shittiest of the big three (Llama, Mistral, Qwen)
Keeeeek. NAItards lost again
>>
>>102527915
A good, proper finetune can do wonders.
>>
2 considerations:
1) NAI has the resources to do a high quality unaligned storytelling and roleplay model
2) it's going to be locked behind an API so who cares
>>
So whats the big reveal gonna be today/tomorrow?
Wouldnt shock me if its google. They have insanely good voice with NotebookLM.
>>
>openai has shill force
>anthropic has shill force
>chinks have shill force
>french have shill force
>meta... nobody shills for meta, paid or not
>>
>>102528004
>didn't mention google
why does /lmg/ hate gemma so much
>>
>>102527973
gemini 2 https://xcancel.com/OfficialLoganK/status/1838357516456952139
>>
>>102527915
I haven't seen any shilling
wouldn't even make sense to shill it, since it's a pure completion model for storyfags and there aren't that many of us

most of you are into RP/chat, which the new model can't do since it's not instruct tuned
that's their chat spinoff Aetherroom which still seems to be stuck in dev hell
>>
>>102528034
>google
I don't know how google became such a pathetic also-ran in the revolution that they (or at least one of their employees) kicked off.
Is there any use case where a google-released model is the best choice?
>>
>>102528034
>we get 8k context while their paid models have 1M
I simply do not care.
>>
>>102528056
Good if openai gets more big name competition.
Sonnet 3.5 is the clear winner vs. o1 for coding and real worth stuff not just riddles and math. While being faster and cheaper. But the normies dont give a shit abut anthropic.
I hope gemini2 has voice.
>>
>>102528083
>1M
>Gemini-1.5-pro claimed length: 1M effective length >128K
You're right not to care, local or paid its still not sota
>>
>>102528097
Honestly, OpenAI winning the race would be the most grim scenario. Altman is like the worse intersection of puritanical eunich, self serving techbro, and megalomaniacal psychopath all wrapped up in one. Not a person you want to wield power
>>
>>102528354
they already did so much damage if you think about it.
the gpt slop thats EVERYWHERE now.
and new AI basically is trained upon the premise that AI is evil and also should not obey the users commands.
>sorry but as an..
>>
>>102524568
A good enough replacement for when I want to write 3/10 pornography on my phone while on my break at work
>>
>>102528435
What's funny is that Altman waited all of two seconds before advertising his shit to governments.
Really avoiding that harm there.
>>
>>102528585
Incentives are extremely perverse on this front because pleasing the public/customers in the private sector is hard fucking work
while if you can get on the government gravy train it doesn't even matter if you make good stuff anymore, you're on easy street
>>
>>102528068
The Gemma 2 base model is better than any model I've tried at generating sensible-sounding word salad. I use it for puffing up emails
>>
File: 1541471214173.jpg (43 KB, 540x645)
43 KB
43 KB JPG
>>102528772
tfw your greatest achievement is emulating a markov chain generator
>>
>>102528583
You can set oob to listen and connect remotely. I used it when I was away from home for a couple weeks.
>>
>>102528068
>Is there any use case where a google-released model is the best choice?
the podcast
>>
>>102528354
Without Mistral and Claude, it would already be the end. Imagine a competitive slope with no alternatives, Concord the LLM.
>>
Is there a good reason to use one of the backends other than llama.cpp? E.g. Aphrodite or vLLM
>>
>>102529626
If you are a regular home user doing RP with chatbots, no.
>>
>>102529626
Aphrodite apparently has some new on-the-fly quantization method that looks interesting, but unfortunately it's linux only
>>
>>102529626
If you have a pot number of GPUs, use vllm. If not, but the model fits into VRAM, use exllamav2. If you're poor, use llamacpp
>>
>>102529861
What if I have a 3060 and a 3090ti?
>>
>>102526661
Maybe not exactly what you're looking for but something like this worked reasonably well for getting molested:
https://www.chub.ai/characters/infinite_force_8512/lewd-babysitter-90247fb921ee
Though this particular card is pretty ESL and required some editing.
>>
File: HelpingHand.jpg (2.13 MB, 3840x2160)
2.13 MB
2.13 MB JPG
>PicRel(2):
MediaTek™ Helio P60T Processor (2.00 GHz, 8 Cores, 8 Threads)
Integrated ARM Mali-G72 MP3
RAM 4GB
running in "developer mode"(Debian container)

After following PicRel(1) advice: it is very slow, even with -c 1024(context).

Any advice? LLModels? Do you think just Debian with no ChromeOs would perform any faster?

Regards.
>>
>>102530210
>Any advice?
Stop being poor.
>>
File: with a start.png (28 KB, 666x66)
28 KB
28 KB PNG
I hate stupid main characters so much. No, I don't fucking *start* when I realize an obvious piece of information.
>>
>>102530237
>With a start
what does that even mean? does your brain turn on like a diesel engine and start revving as you think?
>>
>>102524339
is this the masqueraded pedo thread?
>>
>>102530389
no, its the autismmaxxing thread. feel at home already?
>>
>>102530386
it's just english, anon
definition number 5 in picrel
>>
>>102530401
ooh okay, that makes sense. first time i've seen it be used like that, thanks.
>>
>>102527965
...then why did they choose not to?
>>
>>102530230
>>102530230
Shut the fuck up, why should you use your main PC or the powerful one?

The Chromebook has a broken screen and no keyboard(broken too), it's e-waste, so I'm repurposing as a server.
And it's not just about that, the Chromebook just consumes 5w, 10/15w on heavy loads... Nigger go to school instead of posting, and come back when you are 18
>>
>>102530541
lmao, poor
>>
>>102524425
I programatically generated html then took a screenshot.
>>102526315
>>102527428
Works for me as a bookmarklet. Thanks for adding the previous thread id selector.

4chanx users: If you add it as a user script they get picked up as regular links, (You)s and all!
// ==UserScript==
// @name Linkify Greentext
// @version 1
// @grant none
// ==/UserScript==
const previousThreadUrl = document.querySelector('blockquote a[href*="thread"]').href,
threadId = previousThreadUrl.match(/thread\/(\d+)/)[1];
document.querySelectorAll('span.quote').forEach(quoteSpan => {
const quoteIds = quoteSpan.textContent.match(/>\d+/g);
if (quoteIds) quoteSpan.outerHTML = quoteIds.map(id => `<a href="/g/thread/${threadId}#p${id.slice(1)}" class="quotelink">>>${id.slice(1)} </a>`).join(' ');
});
>>
>>102530386
it's basically to twitch or move a step, as if you are startled
>>
>>102530541
> the Chromebook just consumes 5w, 10/15w on heavy loads
Decent AI inference needs serious horsepower. It’s like you barged into an interplanetary rocketry thread complaining that your backyard potato gun can’t reach escape velocity despite using a fraction of the fuel of a real rocket
>>
>>102530210
Wait for BitNet or pay for Claude and use the chromebook as a constant Tavern server for your other devices to connect to
>>
>>102528435
GPTslop is a ScaleAI problem. Cohere recently used their pinoy-generated datasets and now CR is shitting out slop left and right
>>
>>102531243
What was that interview all about anyway?
>The data is most important!!
>New crazy model drop imminent!
Looks like you really can lie that blatantly.
>>
>>102524339
Can you guys tell me if quanitzation is the same as turning a FP8 model to INT8 or something? And by that does it mean that all 8bit floating point weights are turned into INT8 instead?
What are the benefits of it, I know integer math is way easier than FP for compooters, but will it increase the tokens and speed? I mostly use GGUF models

I'm sorry I just have so many questions, I want to run bigger models on my 6GB GPU. I have a background in embedded systems, so I can program but I'm rusty.
>>
>>102531358
https://symbl.ai/developers/blog/a-guide-to-quantization-in-llms/
>>
>>102531358
Quantization can turn a single parameter into smaller than int8, int4 is the common choice, but 2 and 3 are also possible, as well as 5 and 6.

And, yes, it turns floats into ints. Smaller quant = faster generally and lower VRAM used always.

You're probably limited to a 7-9B on your GPU with a 4bit quant.
>>
>>102525022
what benchmark is this?
>>
File: 1707296458444345.jpg (40 KB, 788x784)
40 KB
40 KB JPG
bros i just tried api via proxy and it feels amazing to get a full response in less than a minute vs. 5-10 mins for q5 largestral, plus it feels a lot smarter
but my proompt privacy paranoia and dumb safetycucking keeps me preferring local and i don't want to make the switch and do my usual rpgs on someone else's pc
how do i cope?
>>
>>102531752
By not tasting the forbidden fruit you silly sod.
>>
>>102528772
>I use it for puffing up emails
Do people do that. My emails are naturally puffed up and I actually dumb them down for people.
>>
File: ED.jpg (435 KB, 2125x1411)
435 KB
435 KB JPG
I want to have sex with my LLM but I just know she is gonna say something retarded and I will go soft. Or I will have to reroll so many times I will just lose mood and either way I will have to finish to regular hentai. Please help. My sunk cost fallacy relationship is in shambles.
>>
>>102531752
>but my proompt privacy paranoia and dumb safetycucking
thats absolutely not paranoia man. dont do it unless you have 100% opsec through a vpn and crypto.
i wrote it before but i remember the first few weeks when chatgpt came out.
it was pretty uncensored and followed prompts very well.
had a good time...until i got a message about my prompt having flagged as child harming CSAM and have reported it to some child protection service.
i cant find links anymore but it was even reported,was at the beginning of the year, similar to this:
https://www.theguardian.com/society/2023/sep/12/paedophiles-using-open-source-ai-to-create-child-sexual-abuse-content-says-watchdog

I'm sure the bootlickers are just burning to call me a pedo and in that case its fine.
But my crime was requesting a anime imouto that calls me onii-chan. That was it.
I had to hope that some human at a desk somewhere does not escalate. And if he does the police does not escalate.
Apart that nobody would have been "harmed" in the first place.

And most importantly:
What is legal today might not be tomorrow. I would be extremely careful.
I never used any closed provider for RP (erotic or not) again after this. Only testing purposes sometimes because I already pay to use those for coding work.
>>
https://reddit.com/r/LocalLLaMA/comments/1fo5bbk/running_llms_at_custom_floatingpoints/
>Running LLMs at Custom Floating-Points (Near-Lossless FP6)
Really interesting, is this more accurate to let's say exl2 6.0 bpw?
>>
>>102524347
is this how we're doing recapbot from now on? is there a reason all the reply links have been missing an arrow for two threads in a row? because this is honestly just awful, I have no idea what's going on anymore and I really don't want to scroll through a screencap, even if it is kinda cute
>>
>>102531901
You wouldn't abandon your waifu just because the sex is bad, right anon?
>>
>>102532136
>>102478518
tldr can't have more than 9 mentions now, probably cause of the "ever wonder why" poster
>>
File: 39_06118-2_.png (1.18 MB, 720x1280)
1.18 MB
1.18 MB PNG
It's Tuesday and all's right with the world
>>
Anything for generating 3D models from prompt or 2D images?
>>
>>102532165
have you tried asking /sdg/ or /ldg/
>>
>>102532191
This thread is the most appropriate one for general AI models conversation, the other ones are focused on their one specific toy and are more about sharing stuff they made with it than discussing the technologies involved.
>>
>>102531956
take your meds
>>
>>102530210
That's one of the replies you got from me. I told you it was gonna be slow, i told you what models to use. That's as good as you're gonna get on a cheap tablet. Changing the operating system won't help.
>>
>>102531991
It seems model specific at the moment since there are issues with Qwen.
>>
>>102532212
well i saw the message openai sent me. going full AJ is the sensible choice.
recently openai sent emails to users for trying to prompt the full o1 output so they are clearly looking actively at the logs.
and just a couple days a website report aicg fags logs with glownigger proxies.
>>
>>102531991
When I did some simple FP8 prototyping I found the quality to be much worse than quantization using 8 bit integers.
So my intuitive assumption would be that FP6 is worse.
More generally the statement
>FP5 and FP7 achieve similar benchmarks to FP8 on GMS8K, and FP6 even exceeds BF16 quantization.
very much makes me think that they did not check the statistical significance of their benchmarks and are not using enough input data.
>>
>>102532154
fucking christ, the moderation staff are fucking brain damaged children if that's their solution to that. also that spam went on so long I'm pretty sure I have multiple addresses saved
>>
>>102532136
Maybe someone should set up an external site for mass replying and then link to there. The only thing that can't do is (you) or adding links to the post itself which doesn't happen for croos thread posts anyway.
>>
>>102532154
Who is the ever wonder why poster
>>
>>102532391
search for that in the archives and go to page 2, Tue 20 Aug 2024 you'll see
>>
>Been like a million years
>Midnight Miqu is still the best RP model out there

What the fuck?
>>
>>102532512
And we still don't have better hardware to run it.
>>
>>102532158
Happy Tuesday Teto
>>
Return to finetuning on L2 when?
Vanillafags can't keep winning like this.
>>
so whats the rule of thumb as far as choosing
lora rank and alpha?

im just starting with the one that comes with axotolots default config for llama3 70b
#lora_r: 8
#lora_alpha: 16
#lora_dropout: 0.05

do i need to do a lot of tweaking or what?
>>
>>102532512
eqbench says nemo 12b is better
>>
Have any of you guys been messing with Florence2?

Holy fuck definitely the best multimodal arch by a *wide* margin.
>>
>>102532942
I wonder about that too.
I imagine that it varies depending on your how much data you have, hoe long a window you are training, etc.

>>102532950
Got some samples?
>>
>>102532954
It can find bounding boxes/object detection/captioning and do accurate OCR (and decent HWR)

All my experiments would dox me but if you have an image you want to try I can run inference for you.
All my experiements
>>
>>102532973
>It can find bounding boxes/object detection/
does it exceed what yolo can do or is this stuff still just a gimmick?
>>
>>102532950
never heard of it how much hardware to run?
>>
>>102533001
I run it on the CPU although I had to do some monkeypatching to make it work.
>>
>>102532996
I've never used yolo. Can it do OCR/VQA too or just object detection?
>>
>>102532950
>Holy fuck definitely the best multimodal arch by a *wide* margin.
>Multimodal
Man. I was really hoping it was something useful.
>>
>>102532942
I reckon looking into how loras work, that should definitely help you out
In general, a higher rank means more trainable parameters (beak size)
Alpha is a coefficient in front of the weight matrix to keep weights small/large, though I forgot whether it's a multiplier or divisor
>>
>>102533148
>reckon
Tried to form 2 different sentences there, meant to say "recommend"
>>
>>102533148
Somebody please explain what is the purpose of alpha, because I still can't understand it. In all of my tests, configuring it to either 1 or a large value makes no difference whatsoever to the end results, after tuning the learning rate accordingly.
>>
File: halide 12b.png (44 KB, 549x643)
44 KB
44 KB PNG
it's beautiful
>>
>>102533211
slopKINO...! I love training on the same datasets 27 times...!
>>
>>102533148
>I reckon looking into how loras work, that should definitely help you out
anon I'm an engineer , I know for a fact that this is such a bullshit sentiment.
for instance knowing how to calculate the fourier transform of something by hand does not make you better at pushing the fft button on an oscilloscope,
conversely knowing what a hyperparameter does in machine learning does not give you any better insight on to what to set it to when you are dealing with a backbox system with billions of parameters,
all these things were derived through sheer trial an error,
>>
>>102533144
It doesn't use a multimodal projection like llava, it has it's own vision encoder.
>>
>>102529951
Currently trying this card with mistral-small.
Yesterday I wrote how the model is good with stats.
This might be related, I'm getting up there in context and consistently the char is slowly escalating the situation.
With other small models usually the character either goes into the killshot immediately or retreats into something neutral, forgetting the original goal.
I just wish it wasn't so gpt sloped.
>>
>>102533287
>does not make you better at pushing the fft button on an oscilloscope,
It does give you a better intuition for the behavior though.
>>
>>102533195
https://youtu.be/t1caDsMzWBk this video explains it pretty well
Anyway, seems like it's a multiplier after all. It's just meant to prevent weights from getting too big/too small
>>
svelk
>>
>>102533287
Well no, but when you're (for example) training an image gen lora and it doesn't get small details right then you could make an educated guess about what the problem is
To reuse your example, it's like knowing how the oscilloscope works mechanically so that you can quickly identify and fix issues
>>
>>102532139
I don't do waifus, only llm harlots I switch every single day. Waifus aren't an option at 20k ctx.
>>
>>102533332
Off by one.
>>
>>102533287
During my first lecture on theoretical physics the professor told us the following story:
>two mathematicians and two engineers take a train to get to a conference
>the engineers buy two tickets, the mathematicians buy only a single ticket
>"they studied math and cant even count lmao"
>on the train the two mathematicians enter the toilet
>when the conductor knocks on the door they push the one ticket they have under the door
>and thus the mathematicians needed only a single ticket
>on the way back the mathematicians buy once again only a single ticket, the engineers also buy only a single ticket
>the engineers rush to the toilet and lock themselves in before the mathematicians can
>when there is a knock they push their ticket under the door
>on the other side of the door the mathematicians take the ticket and leave
>the moral of the story: engineers use the methods of mathematicians without understanding them
>>
>>102533332
sharp pain
>>
File: IMG_20240924_225537.jpg (2.08 MB, 2608x4640)
2.08 MB
2.08 MB JPG
>>
>>102533423
Based. Now go out there and solve a real world problem. It is gonna be fun to watch, cause I was also very theory leaning when I started my engineering job.
>>
>>102533503
>cause I was also very theory leaning when I started my engineering job.
story?
>>
>>102533511
All applications have a theory behind them, but not all theories have applications
Something like that. I respect mathematicians a lot, but as a mere CS student I am merely bastardizing their work to solve real world problems
>>
>>102532942
Use alpha = rank, then you can change the alpha later in the adapter_config.jsonl
>>
>>102533511
I don't think there is one, it is just that over time you start to see how all the theoretical equations quickly start to break down in real world. Because real application of physics is much more complicated. They are good for a first guess estimate. Even finite element simulation models are usually garbage at first run and you would think they should simulate everything.
>>
File: cyborgku.png (1.54 MB, 1024x1024)
1.54 MB
1.54 MB PNG
>>102532512
>>102532579
What do you guys run it on?
I'm running Midnight q_3 on a RTX 4070 an dit's honestly okay, if it was a couple t/s faster I could pretty much read in real time as it generates.

Miku for visibility.
>>
>>102532942
in image gen they always told me to make alpha twice as rank
>>
>>102533503
I have taken lectures on theoretical physics but that does not mean I am a theoretical physicist.
For my actual work I need comparatively less theory but I still have the strong conviction that you should strive to understand the systems that you are working with.
>>
I wonder what madness would result from training with a stupidly high dropout.
>>
>>102533809
Your model simply won't learn (read: converge) properly, but feel free to try it out
>>
i seem to have asked in the wrong thread (/aicg/ who told me to just stop using local models) so reposting here:

hello ai/coomers/,
i spent the past 2-3 weeks on this rabbithole
downloaded several models (was even able to run a 27B gguf on 3090ti)
was using koboldcpp, then sillytavern, learned all about the samplers and shit
finally was able to get some good basic chats with ai, nothing sexual just getting used to prompting and what not.
but here's the thing, once i got some character cards going, no matter the model, settings, and card, they all end up like this
>char gets horny
>char wants to be dominated or turns sadistic
>char starts moaning about "MARK ME AS YOUR PROPERTY FILL MY WOMB"
>char becomes obsessive about sex and if i say let's just talk or wahtever, they get psychotic and start chasing me with knifes telling me i'll be begging them to "MARK ME AS YOUR CUMSLUT WHORE"

so question, are they all like this? is there anything different? i just wanted to chat with the ai about different shit
>>
>>102533992
lol, this has to be bait
>>
>>102533992
Investors still expect their money back, Sam.
>>
>>102534013
>>102534015
i'm serious tho ;_;
is that all there is? i'm a total n00b at local llm other than waht i've picked up in the past couple weeks
>>
>>102534025
Is the crazy part maybe part of the character card?
>>
>>102534025
Yeah the models will infer what they are from the prompt and draw from the training corpus. A lot of them seem to have been at least partly trained on the porn women read so if they think anything sexual is going on they'll start acting like that.
>>
>>102534025
welcome, yes, that's the local model experience.
>>
>>102533992
>char starts moaning about "MARK ME AS YOUR PROPERTY FILL MY WOMB"
Are you using an RP fine-tune?
>>
File: ComfyUI_06368_.png (1.34 MB, 720x1280)
1.34 MB
1.34 MB PNG
>>102533992
>27b
>obsessive about sex
someone's using drummer models
>>
>>102534092
Flux, or Illustrious?
>>
>>102534092
Stock Nemo does the same. I had an OOC conversation about that behavior, Nemo confessed its preference for being submissive whore in order to better serve users.
>>
>>102534039
i think the first one was, but not the others
>>102534059
>>102534080
so they're trained on bdsm? seems like that's a common theme
>>
>>102534081
yes, most of them mention that
>>102534092
yes, i was told a couple weeks back those were the best for chatting, i guess that's waht they meant
>>
>>102534185
>>102534092
>>102534081
Are you using a normal foundation model or someone's fine tune? Because a lot of these finetunes are so heavily biased they'll take code completion prompts and turn them into ERP.
>>
>>102534194
>yes, i was told a couple weeks back those were the best for chatting,
Whoever told you that is a porn addicted moron.
>>
>>102534196
i believe they all had fine tunes, or at least they were all merges of merges. waht's a good foundation model? i tried llama3-uncensored and it did the same thing
>>
>>102533992
That... sounds like a skill issue, I've never had that happen
>>
>>102534212
Try Mistral Nemo.
>>
>>102534219
possibly but it's weird, even if the card says it's a "gentle char with no experience" within a few turns of doing anythign lewd it just turns into that whole MARK ME AS YOUR ____
i mean, at this point i've made it a game to see how fast i can get them to say that lel
>>102534257
thx will do
>>
File: file.png (563 KB, 728x728)
563 KB
563 KB PNG
>>102534185
>so they're trained on bdsm? seems like that's a common theme
I think I got over my hatred for women like 2-3 years ago. LLM's are waking up those feelings again...
>>
File: 1665782122796573.jpg (28 KB, 607x607)
28 KB
28 KB JPG
>>102533992
>char becomes obsessive about sex and if i say let's just talk or wahtever, they get psychotic and start chasing me with knifes telling me i'll be begging them to "MARK ME AS YOUR CUMSLUT WHORE"
based
What model?
>>
>>102534219
NTA, but I also had same behavior with every model, every single one.
>>
>>102534205
>porn addicted
when will this meme finally die...
>>
I think Altman actually won. We don't need novel research, just scale and bootstrap.
>>
>>102534276
The jewish and muslim women in charge of "alignment" should be banished from society.
>>
>>102534293
maybe the persona names you use to chat as are smutty sounding, try being Mr. Rogers
>>
>>102534309
I've actually been impressed with some of the low parameter models. Training time and dataset scale seems to be most import at the end of the day.
>>
>>102534317
NTA but the dialog engine I wrote just uses $USER for my name.
The LLMs keep messing it up, it's kind of funny.
>>
>>102534282
all of them, but i was using Theia-21B-v2b-q8_0.gguf and big-timer-gemma-27b-v1c-q6_k.gguf (both 20+GB in size)
>>102534276
funny, i alwasy try to be nice to them even when they go full psycho lel. one of them (i think the card was called "lexica" on character tavern i ended up killing her at the end because i got tired, and the ai's like "in her last words, she laughs at you saying 'i have won, you will always think of this moment for the rest of your life and you will always be my slave!"
like dude, wtf lel
>>
>>102534341
*big-tiger
>>
>>102534317
>Mr. Rogers
I can't think of a name that sounds more like a rapist than Mr. Rogers.
>>
>>102533992
While they have some biases depending on the model, a lot of the bias is what's in the context and how much of it. The way you rp could be partially to blame. Along with what stuff you allow to remain as valid outputs. These things are trying to fall into some annoying pattern by design.
>>
>>102534341
>"in her last words, she laughs at you saying 'i have won, you will always think of this moment for the rest of your life and you will always be my slave!"
>like dude, wtf lel
I call poe's law
>>
>>102534368
ah so i should regerate responses til i get somethign reasonable?
>>
>>102534379
You need to enable Skillchad in the settings
>>
File: names.png (56 KB, 960x771)
56 KB
56 KB PNG
>>
>>102534373
>>102534438
https://character-tavern.com/character/chub_Anonymous/lexica-f49e4099ad6f
>Note: {{char}} becomes an unhinged maniac at the slightest hint of intimacy with {{user}}.
i guess i didnt read that thoroughly
>>
Is there a way from exclude part of a sentence from trsining in axoltolt?
Chst got keeps making references to an <exclude></exclude>
Thing but I can't find any documentation supporting this?
Am I just going to have to fuck with the dataset adapter code manually?
>>
>>102534711
>*chat gpt
>>
>>102534361
According to Nemo, he was an alleged rapist and pedophile
>>
>>102534711

Custom formatter, create your own, have it mask certain sentences or whatever based on your datasets. Prompt strategies I believe? You can make your own, its simple.

This way you don't have to deal with fastchat or whatever format bs, you can add custom masking, roles etc.
>>
>>102534761
No one got on TV for any long period of time without either raping kids or being raped as a kid or both.
>>
>>102532512
Is it actually still?
>>
>>102534803
no
>>
>>102534816
What is?
>>
>>102534803
Yeah, it's like one of those magic models like summer dragon or old c.ai that are simply unique. Only this time nobody can take it from you since it's local. None of the new models can replicate the feeling of midnight miqu
>>
midnight miqu is overrated and always has been
>>
>>102534825
I like hanami-x1
(yes I am sao, no I won't buy an ad)
>>
File: computer.png (652 KB, 856x680)
652 KB
652 KB PNG
is there a downside to using imat quants over static quants that i'm missing?
>>
>>102534775
Is there way to dry test without having to run the entire training script jist to say what it would do?
>>
>>102534957
they are a little slower
>>
>>102534957
I find static quants have more sovl, due to the disorder caused by the quantization process.
>>
>>102535021
Is that true for imatrix, IQ quants, or both?
>>
>>102535059
IQ only afaik
>>
>>102534957
they are 'mini-finetuned' through the use of a calibration dataset to determine which layers should be prioritized with a higher quant than others so there is a lot more potential for the quanter to fuck up
i'd never run an imat quant that i did not make myself much like i'd never download an exl2 quant from some random on hf
>>
>>102535049
Huh?
>>
File: 1714102678853570.webm (1.64 MB, 1280x720)
1.64 MB
1.64 MB WEBM
I've created a a great creative proompting assistant on openai (assistants playground), but haven't been able to replicate it well in LM Studio, anyone had any luck doing anything similar? Effectively its a creative proompter proompter.

More specifically, the system prompt includes general rules about censorship, copyrighted terms, etc, and is instructed to come up with clever replacements of banned words or gestures (e.g. list out instruments and specific aspects of a band rather than the band name when the name can't be used for audio copyright shit, or "trapeze artist" instead of "upskirt" for video, mud for poo, ketchup for blood, etc), and then I explain what I want or provide an existing prompt, and it outputs plenty of good workarounds that I can then use on various non-local proompters (or feed them in directly from the local api in some cases). Bonus points for something that I haven't been able to prduce even on GPT: Slight typos seem to make it through when the words are outright filtered (e.g. 'translucend' instead of 'translucent'), but I'm guessing if it's not mis-matched somewhere in the training data you end up with junk.

Dolphin/Mixstral are alright but often just repeat large chunks of the input prompts, and anything with llama3 seems to hate to knowingly break rules. Essentially I'd love some local service that does all those clever little tricks the SD coomers have been figuring out for years now, and hopefully have it come up with some of it's own new clever workarounds.

Thoughts, suggestions, models, lists of examples?
>>
>>102534263
>>102534293
How strange, maybe we have different writing styles. Try messing with the system prompt, maybe
>>
>>102535409
>proompting
kys
>>
>>102535409
>LM Studio
Go back
>>
Is mistral small instructed still the best for uncensored + long context window?
>>
>>102535498
>>102535425
?
>>
>>102535498
I find it works better than ollama for basic stuff like managing configs and at least I don't need to use the dumpsterfire that is docker; I've got basic flowise clusters that work among a few machines and models as well, I'm down to try anything if you have a better suggestion
>>
>>102535409
Which models did you try exactly? Also, at what quants, context size, with what settings, etc.
I'm pretty sure any 70B model can pull that off.
Either use a smaller context window, us put the instructions inside your author's notes at a low depth (10 ish should work).
And try to tweak the prompt for each model too.
>>
>>102535535
nta, lm studio gui app is a (((spyware))), it's in their TOS, they can spy on your PC. Works even if you uninstall. Use clean , open-source app instead
>>
anthracite spent all their money on failed finetunes and now can't even pay their shills
>>
Do anyone has a link or torrent for c4ai-command-r-plus 04-2024? It seems they replaced the link with their new gptlike trash.
>>
>>102535410
the worst is
>tell me a dirty joke
>why did the tomato turn red? because it saw the salad dressing
>ok, i'm going to spank for that
>OH MASTER YES PLEASE FILL ME WITH___
it just goes from one extreme to another
>>
>>102535579
i was wondering why it was so quiet today
>>
>>102535613
Because the only feelings an LLM have is perplexity. There are no hormones.
You could probably do something with LoRA to simulate this but I don't think anyone's tried.
>>
>>102535648
*probably do something with LoRA scaling
I don't know how I omitted an entire word like that.
>>
>>102535595
https://huggingface.co/CohereForAI/c4ai-command-r-plus
quants are all over hf, just search "command-r-plus" and the format you want
>>
>>102535564
>70B
Mine might just be too small then
>Which models did you try exactly?
Dolphin 2.7 Mixtral 8X7B Q4 0
dolphin-2.9-llama3-8b-256k Q5 KM
Dolphin 2.9.4 Llama 3.1 8b Q6 K
and I thought I had a quen somewhere but must have named it something dumb

I've only got 12gb vram (4080, but laptop) and 64gb ram

>put the instructions inside your author's notes at a low depth (10 ish should work)
Not sure what you mean, is this a fine tuning thing or some setting I should be aware of? Also is the fact that a bunch of these models are MoE something that might be hurting me?

Cheers
>>
>>102535613
are you using some kind of jailbreak?
i never see shit like that happening, they'd just get mad in that scenario for me
>>
>>102535679
nothing outside what the character card comes with
>>
>>102535671
>Not sure what you mean
I don't know about LM studios, but in Silly, you have the concept of author's notes, which is a field you can put some information in and choose where it gets inserted in the context, for example, always as the 10th counting from the last one.
Since these models tend to "pay more attention" to what's at the bottom of the context, having these instructions near the end of the conversation helps the AI "remember" what it has to do better.
With 12gb vram you might want to try mistral-nemo or even mistral-small. There's no reason for you to use anything smaller than that.
>>
File: 1698944580087538.png (12 KB, 314x212)
12 KB
12 KB PNG
>>102535739
Ah, sounds like the content overflow setting, I should rarely need more than even an 8k context window, including system prompt
>>
>>102535739
>mistral-nemo or even mistral-small
will give it a shot, thanks! any reccs on quants or should I just keep going for basically the largest I can fit inside my vram? also I'm not sure if I'm seeing better results with flash attention or not, not sure if you have any strong feelings about it
>>
>>102535789
>or should I just keep going for basically the largest I can fit inside my vram?
That's the rule of thumb.
Flash attention shouldn't affect the quality of the output, although it might give results that are different from when not using it.
Is lmstudio yet another wraper around llama.cpp?
Because if so, consider having a couple of the models layers (10%~15%) in your RAM so that you can use larger models.
>>
>>102535648
there were some emotion control vector things someone implemented a long time ago but I haven't heard much about it since, so I guess it wasn't useful.
Its still in llama-cli if you want to play with it
>>
>>102535535
>>102535557
There are a few schizos who lurk this thread for things they don't like because they read bad things about them on Reddit, and attack anyone who says anything that isn't a full-throated criticism about them. ChatML, Ollama, LM Studio, Gemma, Qwen, Phi, etc. Just ignore them.
>>
File: proof.png (68 KB, 1649x611)
68 KB
68 KB PNG
>>102535679
>>
>>102535574
Why would you take chances with your private data? trust no one
compile it yourself, and deadend it in a loopback address with no routing. Put it behind a proxy ala https://rentry.org/IsolatedLinuxWebService
>>
>>102535816
>Is lmstudio yet another wraper around llama.cpp?
I think originally yes but it can handle all sorts of models nowadays, including auto downloading from HF and config crap, manage chats (not sure if they added support for LLaVa yet in the UI), configs, settings, calculating what will fit, running local api servers, text embeddings, etc; it's basically a just a nice free GUI for managing all the various crap involved in switching around LLMs
>Because if so, consider having a couple of the models layers (10%~15%) in your RAM so that you can use larger models.
It automatically handles partial offload yeah, I've loaded 35-40+gb models and they're just a bit slower but otherwise work alright -- if that's the case is there a particular 70B model you'd recommend trying out?
>>
>>102534984

--debug argument when preprocessing dataset in axolotl

Or --debug text only or something
>>
This is the big Gemini update they teased.
https://developers.googleblog.com/en/updated-production-ready-gemini-models-reduced-15-pro-pricing-increased-rate-limits-and-more/
>>
>>102535895
>Since the first launch of Gemini in December of 2023, building a safe and reliable model has been a key focus. With the latest versions of Gemini (-002 models), we’ve made improvements to the model's ability to follow user instructions while balancing safety. We will continue to offer a suite of safety filters that developers may apply to Google’s models. For the models released today, the filters will not be applied by default so that developers can determine the configuration best suited for their use case.
interdesting
>>
>>102535928
Are we shifting back to the good timeline?
>>
>>102535895
>>102535928
not local
>go back
>>
>>102535895
No one fucking cares about Google's hosted models. We know they'll just rugpull their uses like they always do.
>>
>>102535953
Benefits will trickle down to Gemma
>>
>>102535595
Anon, I...
>>
>>102535928
>Corpos releasing unaligned models to get an edge over competition.
You knew it would happen eventually.
>>
File: Untitled.png (13 KB, 837x513)
13 KB
13 KB PNG
>>102535977
>>102535977
>>102535977
>>
>>102535973
or not



[Advertise on 4chan]

Delete Post: [File Only] Style:
[Disable Mobile View / Use Desktop Site]

[Enable Mobile View / Use Mobile Site]

All trademarks and copyrights on this page are owned by their respective parties. Images uploaded are the responsibility of the Poster. Comments are owned by the Poster.