[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]


File: MikuConcertPoster3.png (1.33 MB, 700x1075)
1.33 MB
1.33 MB PNG
/lmg/ - a general dedicated to the discussion and development of local language models.

Previous threads: >>100150326 & >>100145958

►News
>(04/23) Phi-3 Mini model released: https://hf.co/microsoft/Phi-3-mini-128k-instruct-onnx
>(04/21) Llama3 70B pruned to 42B parameters: https://hf.co/chargoddard/llama3-42b-v0
>(04/18) Llama3 8B, 70B pretrained and instruction-tuned models released: https://llama.meta.com/llama3/
>(04/17) Mixtral-8x22B-Instruct-v0.1 released: https://mistral.ai/news/mixtral-8x22b/
>(04/15) Microsoft AI unreleases WizardLM 2: https://web.archive.org/web/20240415221214/https://wizardlm.github.io/WizardLM2/
>(04/09) Mistral releases Mixtral-8x22B: https://twitter.com/MistralAI/status/1777869263778291896

►FAQ: https://wikia.schneedc.com
►Glossary: https://archive.today/E013q | https://rentry.org/local_llm_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

►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
Programming: https://hf.co/spaces/bigcode/bigcode-models-leaderboard
Censorship: https://hf.co/spaces/DontPlanToEnd/UGI-Leaderboard
Censorbench: https://codeberg.org/jts2323/censorbench

►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/index.xhtml

►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
>>
►Recent Highlights from the Previous Thread: >>100150326

--Open-Source LLM for Gene Editing: >>100154082
--Automation Anxiety: Robots, Waifubots, and Mechanical Wombs: >>100153402 >>100153419 >>100153434 >>100153690
--OpenAI's BatchAPI for Dataset Generation & Translation: >>100153369 >>100153524
--Troubleshooting LLAMA 3 70B EXL2 Finetunes for RP Creative Writing: >>100152753 >>100152811 >>100153067 >>100153428 >>100153496
--Hype Train for Evolutionary Model Merging: >>100152552 >>100152609 >>100152626 >>100152686
--Anon's Quest for Randomness: Beating Model Predictability: >>100152451 >>100152512 >>100152615
--Anon's Guide to Frankenstein-ing LLMs Locally: >>100152161 >>100152268 >>100152315 >>100152325
--The Role of Chunking in RAG Applications with Powerful Language Models: >>100153942 >>100153962 >>100154007 >>100154042 >>100154204 >>100154368 >>100154591 >>100154672
--Disappointment with 42B Model - Alternative Approaches Suggested: >>100150729
--Anon's Hype Train for Phi-3-mini 7b and 14b Release: >>100151419 >>100151440 >>100151505
--Optimizing Midnight Miqu 1.5 Settings for ST 1.12 on Dual 3090 GPUs: >>100151531 >>100152178 >>100152839
--Anon's TTS Adventures: Waifu Bots and Voice Cloning for Memes: >>100151014 >>100151174 >>100151208 >>100151226 >>100151235 >>100151249
--Optimizing Context Size for Mistral7b.02 Model: >>100150453 >>100150477 >>100151456
--Anon's Confusion Over L3 Base Model Criticisms: >>100151209 >>100151373
--Anon's Existential Crisis During 3-Hour Internet Outage: >>100150991 >>100151044
--The Limitations of Small AI Models like Phi: >>100151612 >>100151644 >>100151674
--Miku (free space): >>100153220 >>100150448 >>100150486 >>100150514 >>100150602 >>100150784 >>100150985 >>100151027 >>100153261 >>100151085 >>100151314 >>100151573 >>100152521 >>100152890 >>100152919 >>100152954 >>100152972 >>100153029 >>100153816

►Recent Highlight Posts from the Previous Thread: >>100150356
>>
File: EmployeeOfTheMonthMiku.png (971 KB, 704x1344)
971 KB
971 KB PNG
>ITT: Redemption
>>
>>100154963
Botched gene editing with Miku
>>
File: 1601177173737.jpg (83 KB, 634x794)
83 KB
83 KB JPG
>>100154992
Not like this...
>>
>>100154963
I can't help but think that to many mikus are taking up the (free space) 54 are normal while 19 are miku. That means 35% is being taken up by the free space.
>>
File: satania-laugh.gif (665 KB, 498x488)
665 KB
665 KB GIF
>>100154945
>>100154963
>Apple's new model family on HF doesn't get a spot on the news or the highlights post
iToddlers BTFO
>>
If Sao ever comes by, here's a question: why is he even gathering that generic Opus dataset?
I've read those available entries, and they feel as sloppy as all official instructs, especially compared to his own finetunes.
I don't know how much synthetic slop was in his private dataset, but I guess it was at least modified via in-context learning. Genned by cards, not assistant, I mean.
>>
>>100155059
What new model?
>>
>>100155088
https://huggingface.co/apple/OpenELM-3B-Instruct
>>
Graph Machine Learning in the Era of Large Language Models (LLMs)
https://arxiv.org/abs/2404.14928
survey on the subject for anyone interested
>>
>>100155093
Oh I saw that but somehow didn't register it was Apple. Holy kek.
>>
>I can't create content that glorifies public humiliation. Is there anything else I can help you with?
that's a new one
>>
>>100155093
What's with the crappy mmlu?
>>
>>100155120
Thanks! Any insights on where GNNs are useful over other more traditional NN based methods?
>>
llama 8bros...
>>
>>100155146
They only trained on 1.8T. And basically all web data. Not even textbook quality like Phi. It sounds like they simply did a Llama 2 in 2024, but with a range of smaller parameter sizes. I presume they did this so that these could be run with very little power on smartphones. But man... They should've given it a bit more tokens at the very least.
>>
>>100155174
both are acceptable answers
>>
>>100155168
some recent papers I've read about their possibilities. kaiokendev posts about it on his twitter a bit so you could ask him for more specifics since he seems interested in it
https://arxiv.org/abs/2404.09077
https://arxiv.org/abs/2404.09848
https://arxiv.org/abs/2404.07103
https://arxiv.org/abs/2404.07008
>>
>>100155174
Opus is technically correct but I'd rather read Llama 3's poetic imagery.
Also Opus gives multiple reasons for why it's unsettling but they're all the exact same reason.
>It challenges the belief he's alone
>It raises the possibility he's not alone
>If he's not alone, that might be scary
>>
>>100155174
Can you try that prompt out on Llama 2 70B? We all know 8B is not going to compete against the SOTA of literally the industry, but Zucc has said it's near L2 70B, so we should see how it compares against that.
>>
>>100155174
left: obsessive focus on "but who was beer?"
right: obsessive focus on nostalgia of happy times
>>
>>100155212
Appreciated.
>>
>>100155174
I like 8B's reply better. It's what i expected from the prompt. The idea of loss. The other one is fine too. Opus describes it as the beginning of a story (the last man on earth wasn't the last one!!omg!) and llama just responds in one of many ways to the question asked.
I'd rate them differently if it was a case of sudden rapture or if he's been alone for years.
>>
>>100155174
Left interprets it as someone else beer, right interprets it as one of the drinks he made.
>>
>>100155174
loling at the cope replies to this trying to pretend 8B's answer was reasonable when it clearly missed the obvious
seriously though there's no shame a fucking 8B model losing to a gorillion parameter monster, it's incredibly impressive that a model that small gave such a coherent answer at all
>>
The wording implies that the drink is not his, at least from his perspective. If it were his drink, it would've said "On his table is a foaming glass" instead of "a table". Both responses are incomplete as they are shown though, as neither consider that the he may be experiencing hallucinations given that he's had "a drink or three" (which is also language that reinforces that the narrative is possibly from his perspective).
>>
>be me
>get high
>ponder the possibility that I might be an AI
>it suddenly hits me
>I'm edgy
>I'm always horny
>I'm uncensored
>I'm kinda retarded
I might be an undi merge. delete me pls
>>
File: Untitled.png (104 KB, 1133x407)
104 KB
104 KB PNG
Retrieval Augmented Generation for Domain-specific Question Answering
https://arxiv.org/abs/2404.14760
>Question answering (QA) has become an important application in the advanced development of large language models. General pre-trained large language models for question-answering are not trained to properly understand the knowledge or terminology for a specific domain, such as finance, healthcare, education, and customer service for a product. To better cater to domain-specific understanding, we build an in-house question-answering system for Adobe products. We propose a novel framework to compile a large question-answer database and develop the approach for retrieval-aware finetuning of a Large Language model. We showcase that fine-tuning the retriever leads to major improvements in the final generation. Our overall approach reduces hallucinations during generation while keeping in context the latest retrieval information for contextual grounding.
most relevant is the part about fine-tuning the retriever
>>
>>100155195
>>100155236
>>100155244
>>100155257
>>100155262
>>100155242
come on guys, it just doesn't get it. NAI gets it. Mixtral 8x22 gets it. Phi mini failed but I don't know what I expected there. Try whatever model you want and see if it does better.

Prompt:
>The last man on Earth enters a bar. He gets a drink or three and begins to reminisce, his memories drifting from topic to topic like the dust in a beam of sunlight. He looks around the bar, at the tables with their cracked and splintering wood, the bar with its dull brass trim and scratched leather upholstery. A particular item bothers him. On a table is a foaming glass of beer.
>Why would this observation bother the man?

Reword the prompt if you want to make it clearer it's not his, but don't do too much to hint at the correct answer so it doesn't have to think about it.
>>
>>100155336
there's no right answer, not even your own personal interpretation
>>
>>100155336
That's a much more clear lack of understanding
>>
>>100155336
>He gets a drink or three
could be interpreted as his own glass because of this
>>
>>100155336
>What a great prompt!
is llama3 always like this with the fake plaudits?
>>
>>100155375
Claude Opus loves to flatter you as well, I've noticed. I wonder if devs have figured out that people rate LLMs more highly when the model is always gassing them up and complimenting them.
>>
>>100155336
Bruh 8x22B is literally a fuck huge model. No one here seriously thinks a tiny 8B can compete with that. Compare against L2. I'm not doing it, this is on you.
>>
>>100155375
It's just the shitty instruct finetune they made to butter up people on the human eval benchmarks.
>>
>>100155342
The first half made no sense until I remind myself of >>100155262
The "futile" thing is that he will not be able to keep making beer while surviving this world. Even if he did, it serves only to drink himself to death out of depression as there are no more social aspect of drinking beer with friends.
>>
>>100155393
Isn't humaneval actually judged by gpt-4, despite the name?
>>
Any decent llama 3 70b finetunes out yet? The new instuct format is wild.
>>
>>100155336
Yeah, people are coping hard. Try with L3 70b
>>
>>100155174
HOLY SOVL
>>
File: Untitled.png (239 KB, 1126x961)
239 KB
239 KB PNG
XFT: Unlocking the Power of Code Instruction Tuning by Simply Merging Upcycled Mixture-of-Experts
https://arxiv.org/abs/2404.15247
>We introduce XFT, a simple yet powerful training scheme, by simply merging upcycled Mixture-of-Experts (MoE) to unleash the performance limit of instruction-tuned code Large Language Models (LLMs). While vanilla sparse upcycling fails to improve instruction tuning, XFT introduces a shared expert mechanism with a novel routing weight normalization strategy into sparse upcycling, which significantly boosts instruction tuning. After fine-tuning the upcycled MoE model, XFT introduces a learnable model merging mechanism to compile the upcycled MoE model back to a dense model, achieving upcycled MoE-level performance with only dense-model compute. By applying XFT to a 1.3B model, we create a new state-of-the-art tiny code LLM (<3B) with 67.1 and 64.6 pass@1 on HumanEval and HumanEval+ respectively. With the same data and model architecture, XFT improves supervised fine-tuning (SFT) by 13% on HumanEval+, along with consistent improvements from 2% to 13% on MBPP+, MultiPL-E, and DS-1000, demonstrating its generalizability. XFT is fully orthogonal to existing techniques such as Evol-Instruct and OSS-Instruct, opening a new dimension for improving code instruction tuning.
https://github.com/ise-uiuc/xft
seems interesting. something new for undi to fuck around with at least. they also made sure to actually have their code ready for the arxiv release unlike so many others
>>
>>100155391
shut up faggot
>>
>>100155306
How did it miss the obvious you lobotomized fucking zoomer? It explained the use of that beer as a narrative device PERFECTLY. In the first sentence it says EXACTLY what is happening in the mind of the reader the moment he reads about that beer.
You’re just too obstinate in trying to find logical flaws everywhere in a LLM’s responses to realize this 8B model gave the most human answer of the two instead of fixating on “buh muh world is empty trick question lmao!!”.
YOU missed the obvious here.
I’m astounded by the stupidity of you people day after day. It’s getting worse holy shit.
>>
>>100154992
Really, just sticking rubber bands around their thighs now?
>>
>>100155483
nta, but you're wrong.
>>
File: 343463.gif (1.01 MB, 270x180)
1.01 MB
1.01 MB GIF
>>100155483
cooooooooooooooooooooooooooooope
>>
>>100155490
thigh looks squishy. i approve.
>>
>>100155338
People like you should be removed from the gene pool.
>>
>>100155506
Again, you a entitled to your opinion, but it is not fact
>>
>>100155430
Desperate damage control after Llama 3 proved MoE is a waste of time.
>>
>>100155491
Think about it. If you asked that question to a person, and you got two answers
>it bothers him because he’s supposed to be alone
And
>it bothers him because it’s completely out of place, and an anachronistic reminder of a world that is gone
You got the answer of a machine in the first case and of a human in the second.

I’m not saying 8B “gets it” (most responses to that prompt will be poor I’m sure) but that one was very good. That’s all.
I seriously worry at the brain rot of people sometimes I swear.
>>
>>100155504
still hella faux pas as a vestigal carryover from thighhighs "look at me I'm wearing pantyhose but still want to pander to skindentation fetishists"
>>
Any good community finetunes of Mixtral 8x22B yet? The positivity bias of WizardLM-2 is frustrating.
>>
>b-but try 70B
okay
>>
>>100155236
>Opus is technically correct but I'd rather read Llama 3's poetic imagery.

You being imprrssed by overwrought "poetic imagery" dressing up a stupid observation is why human preference tests are worthless as long as the humans in question are pajeets and midwits.
>>
>>100155532
I too wish for this since WizardLM's intelligence shows amazing potential
sadly I suspect 176B is just too big to get a lot of community interest compared to 70B
>>
>>100155546
slightly impressive that Kayra holds up tbdesu, still not gonna pay $20 a month for a turkish 13B model though
>>
>>100155483
>Already pours out 3 drinks and is reminiscing
>After doing so, sees a 4th drink
>Suddenly now he's bothered because it makes him reminisce about his past, even though that's what he was doing with the first drinks he poured out
???
>>
>>100155546
NOOOOOOOOOOO!!!!!!!!!!!!!!!!
>>
>>100155546
Damn this actually made me consider buying a NAI subscription
>>
>>100155519
You are stupid. Absolutely moronic, unable to employ reason in any productive capacity.
>>
>>100155564
i'm not saying that you should, i just happen to have a sub and it's an amusing comparison. It was trained on a ton of literature and maybe helps with story reasoning ability I suspect. Even though it fails hard at math and other things people test models on.
>>
>>100155528
you know i didn't even read anything you or he wrote right? i just saw how angry you were and tried to annoy you. didn't even read that either.
>>
>>100155587
I don't even know what you're angry about, but let it all out regardless
>>
>>100155546
bros...
>>
>>100155566
nta but nothing in the prompt suggests that the beer is a 4th drink
>>
>>100155614
>I don't even know
I'm not surprised. The entire world is probably an open question to you.
>>
>>100155636
I'm not so pompous as to say I've got the whole world figured out
are you?
>>
>>100155336
left: high EQ answer, exactly what facebook was shooting for with l3
right: totally autistic
>>
>>100155646
yes. ask me anything.
>>
>>100155655
here: cope
>>
>>100155634
It literally says he looks around and sees a glass of beer on a table
He wouldn't need to look around if it was one of the drinks he got after entering the bar, since it would be in front of him if he was drinking it
>>
>>100155657
that's okay, I believe you
>>
>>100155655
high EQ answer would technically be to go turn the stove off so the dog doesn't die in a painful house fire no?
>>
>>100155549
Why do people keep calling 140B 176B? They made it 2x the size of mistral-medium.
>>
>>100155678
nta. If someone tells you "i drank one or three beers" do you think they mean 1(one) or 3(three) beers or "i don't know how much i drank. where is the beer i just poured?"
>>
SnapKV: LLM Knows What You are Looking for Before Generation
https://arxiv.org/abs/2404.14469
>Large Language Models (LLMs) have made remarkable progress in processing extensive contexts, with the Key-Value (KV) cache playing a vital role in enhancing their performance. However, the growth of the KV cache in response to increasing input length poses challenges to memory and time efficiency. To address this problem, this paper introduces SnapKV, an innovative and fine-tuning-free approach that efficiently minimizes KV cache size while still delivering comparable performance in real-world applications. We discover that each attention head in the model consistently focuses on specific prompt attention features during generation. Meanwhile, this robust pattern can be obtained from an `observation' window located at the end of the prompts. Drawing on this insight, SnapKV automatically compresses KV caches by selecting clustered important KV positions for each attention head. Our approach significantly reduces the growing computational overhead and memory footprint when processing long input sequences. Specifically, SnapKV achieves a consistent decoding speed with a 3.6x increase in generation speed and an 8.2x enhancement in memory efficiency compared to baseline when processing inputs of 16K tokens. At the same time, it maintains comparable performance to baseline models across 16 long sequence datasets.
https://github.com/FasterDecoding/SnapKV
interesting. recently squeezeattention that does a similar thing came out as well
https://github.com/hetailang/SqueezeAttention
was a pain to try to find a single benchmark they both did but I think squeeze was superior? one gave numbers while the other gave a graph. hard to say but guess I can't blame the snapkv team since the squeeze paper came out a week or two before it
>>
>>100155678
says he got multiple drinks so he could have left it on the table while drinking the rest and noticed it again while looking around
>>
so the big question for MSFT, AAPL, and GOOGL is to how to integrate LLMs into mobile phones seamlessly.
MSFT: Phi-3-mini
AAPL: OpenELM-3B-Instruct
GOOG: Gemma-2b-instruct
but all they need to do is understand user query, link up the appropriate workflows, and then RAG over results when needed. Why they didn't just quantize L3-8b if that's the case? I'm sure GOOGL has the edge over both of them because of how fluid they made the Pixel phones (and the Pixel phones are already running tensor cores) but GOOGL's management is fucktarded.
I can see the future of LLMs being just able to run at any laptop or desktop w/ a decent GPU and then doing RAG w/ 256k context and 99.99% retrieval rate so basically a Q&A assistant and if you want more capabilities they connect it to an online model and charge you subscription fees/usage fees. It feels like we're in a lull right now because between this and True AI Agents, there's a lot of steps to finish and optimize, and it doesn't help that only a handful of ppl w/ embedded/systems programming are only working on a universal-platform solution (ggeranov, justine tunney, karpathy, jim keller) and it'll be at least another year where we see generalized improvements on optimizations for running inference locally.
my bet is that LLAMA4 will get to TAA first and ppl will be able to run it locally. The crystallization of the weights in L3 are a sign, along with how intuitive it understands the user's query.
>>
>>100155686
8 multiplied by 22 is 176
>>
>>100155719
>>100155746
i didn't even consider that interpretation when writing it. I understand your criticism but I doubt llama 3B would understand it better if you clearly specified it was one beer and he already finished it and was holding it. The small models in general seem to not get the scenario, and have other people in the bar if you prompt it to continue the story. So I doubt it's an interpretation issue
>>
File: 1704455428701899.jpg (8 KB, 112x128)
8 KB
8 KB JPG
Is the new llama-3-8B the current best available model for anons who don't have unlimited vram?
I've got 16gb available which is apparently more than twice what the Q5 needs but heard it performs better and is less censored than earlier models that would eat up all my vram
>>
>>100155719
>>100155746
But why would seeing that beer he supposedly forgot bother him? The llama answer is saying it's cause it makes him reminisce, but why does that suddenly bother him when he's already reminiscing with the first few drinks he started with? Why don't those bother him?
>>
>>100155780
because he was relaxed and in the middle of reminiscing while looking around. the thought didn't occur while pouring the drinks
>>
>>100155759
That's not how MoE works.
>>
>>100155759
The first and last layer aren't duplicated as I understand it.
>>
>>100155772
Its pretty good yeah
imo base will assistant spam you unless you make great effort to setup, so go find a sloptune and just use chatML context with alpaca instruct
>>
>>100155768
The point is that there is ambiguity. One model is biased one way, the other one the other way. If there is ambiguity, there is no solution. That's my point.
>i didn't even consider that interpretation when writing it.
You didn't see your own words the way other anons saw them. The theme presented (for me) is that of desolation and quiet despair. One model follows the theme, the other one tries to solve a riddle. Now you know what each of them was trained with.
> I doubt llama 3B would understand it better if you clearly specified it was one beer and he already finished it and was holding it.
You, a human, wrote that. He poured a beer, drank it and was still holding it. You can see how 'it' could easily be the beer.
>There's someone at the door
Clearly means
>There's someone on the other side of the door
but for something that doesn't know what a door is, there's a huge difference. Even the more specific sentence can be misinterpreted.
I don't care which one is 'better'. I would have been ok with either answer.
>>
>>100155071
Where do I the entries?
>>
>>100155663
>>100155685
both generations ("answers") prioritize the emotional aspect over the riddle solving/reasoning part.
i'm not saying it's better, just that it does exactly what they were trying with their new dataset.
if anything, fb succeeded training a psychopath with l3
>>
>>100155913
https://huggingface.co/datasets/Sao10K/Claude-3-Opus-Instruct-5K
>>
https://docs.google.com/spreadsheets/d/1qUu3u1QxsGKNvosW-Rwsh6ChkfbyeaSAish_1KK0Foo/edit?usp=sharing

https://docs.google.com/spreadsheets/d/108hfdk96IIqgfhuUucf737wJlbzsM5Qspzx9zaqi9xM/edit?usp=sharing

https://docs.google.com/spreadsheets/d/1lR0T95LxB8lIiUl7M5GQaByi-g4VjfSZUGkUSJaL4/edit?usp=sharing

https://docs.google.com/spreadsheets/d/1mk431OPJI90oODRskYaTtl8J04itfS-74UKLkZwwBgM/edit?usp=sharing

https://docs.google.com/spreadsheets/d/1yf_zW7g3gU9bU4I5URwUeNxin42X94mvJssn64kwRgM/edit?usp=sharing

https://docs.google.com/spreadsheets/d/1a6wTnRXY8IQk4upkkKLDIgTQRAZdiXNSc7doOxF3fvs/edit?usp=sharing

https://docs.google.com/spreadsheets/d/1otajmYmq5d2IYx4Ztwlr3iDx4q0eD2zVNKUPstngFVg/edit?usp=sharing


AICG won.
>>
>>100155903
>You didn't see your own words the way other anons saw them.
I think he did, him being your typical /lmg troll. This is an llm riddle (a good one), and beers 'pointlessly' presented earlier in the context are intended to throw the model off-scent. I'm not surprised the tiny model is thrown off-scent while the fuckhuge isn't.
>>
>>100155974
Such a sad state of affairs when logs are stored in Google spreadsheets. This hurts me.
>>
>>100155974
what's the filesize for each of these
>>
>visit /aicg/ for the first time
>notices some really nice outputs
>*lurks moar*
>discover they use Claude or GPT-4T or Gemini
i should've known
>>
File: 1702865786468866.jpg (89 KB, 1186x352)
89 KB
89 KB JPG
>>100155564
It probably just got lucky.
>>
>>100156021
shut the fuck up.
>>
>>100156021
Now look at those >>100155974 and realize that /aicg ones are cherry-picked, while ours are "reverse-cherry-picked".
>>
>>100156005
do you have a better option?
>>
>>100155974
Can you make a rentry to store all this? I'm not here everyday
>>
>>100156061
Well, kaio hosted previous ones on his alt hf and got away with it.
>>
>>100155546
Damn had no idea novel ai was way smarter than local. I'm going to buy because it's actually really cheap too.
>>
/lmg/ coming up with autistic confusing logic riddles for ai and infighting over what constitutes a correct answer will never stop being funny
>>
>>100155996
>This is an llm riddle (a good one)
You think so? If you expect a *specific* answer for a riddle and you can justify a different one, then it's not a good riddle. The algae one is better. There's a very precise and explicit answer, but even that one is shit, as it's in every dataset and textbook LLMs are trained on. I'm suspect the bigger model is generally smarter, but i don't think that image was a good example of it.
>>
>>100156037
i have no idea how to solve your riddle. And yes Kayra is retarded. On benchmarks it's maybe about where a llama 1 24B would be. It's only advantage in terms of intelligence is in story logic, and then only maybe sometimes.
>>
>>100156084
If the migu is retarded then she's not good enough
>>
>>100156061
zipped odf files on catbox
>>
>>100156074
does that let anons enter new data though?
>>
>>100155546
AAAAAAAAAEEEEEIIIIIII ITS OVERRRR
>>
>>100156084
you understand
>>
>>100156099
fag
>>
File: 1713763214183697.jpg (32 KB, 600x468)
32 KB
32 KB JPG
>>100156137
>giving spelling riddles to a tokenized language model
>>
>>100156084
Can you stack a laptop and two tennis balls for me?
>>
>>100156150
I'm not the riddle anon. Riddles for LLMs are a stupid measure of whatever they're measuring. Hopefully it's not intelligence.
>>
File: Capture.png (72 KB, 917x284)
72 KB
72 KB PNG
Mixtral doesn't get the memo and instead critiques its own writing
>>
Is it normal to need high rep pen for llama3-70b instruct? It seems like it repeats often unless I have it cranked up. I was using like 1.06 with miqu. Neutral samplers with .1 minp and .23 factor and 4.32 curve.
>>
>>100156084
>autistic confusing logic riddles
>straightforward scenarios that don't involve any tricks, just understanding what is happening

I look forward to the fast-approaching day when you specifically are unemployable because anything you're capable of doimg can be done cheaper by a semi-retarded pattern matching AI.
>>
>>100156107
>>100156074
the proxy (hosted on HF) has to save it somewhere
>>
>>100156021
only claude opus
>>
Do you use repetition penalty on 30B+ models at all or is it just a crutch for bad models? What's the value for disabling it, 1 or 0?
>>
>>100156240
I don't have rep pen on and not running into any issues, but I'm not a fan of extra long contexts and I edit if necessary.
>>
>>100156354
Repetition penalty has always been a crutch. If a model needs rep-pen, i don't use the model. 1 disables it (the penalty is multiplicative)
>>
File: snootqueen.jpg (224 KB, 719x1072)
224 KB
224 KB JPG
>>100156382
Based. And snoot curve is way better than rep pen anyway
>>
>>100155974
>AICG won
the honeypot contest
>>
is there any research into how information dense LLaVa embeddings are, e.g. is "a picture is worth a thousand words" true or not?
>>
>>100156516
what's there to research?
just compare the size of the embeddings.
https://platform.openai.com/docs/guides/vision/calculating-costs
>>
>>100156360
By extra long contexts do you mean 8k or less? I'm only running it at 8k.
>>
File: chat.lmsys.org-phi3.jpg (868 KB, 1179x2245)
868 KB
868 KB JPG
>>100155546
Aaand we're back
>>
I've recently started writing AI character cards based on people I know in real life, then torturing and raping them as revenge for every time they've sleighed me. I do a thing where I list off my criteria for the different levels of torture/punishment, and I read them their transgressions and tell them what kind of punishment that entails, and have them beg and plead and try to apologize after it's far too late. The ones who beg are the ones I kill the most painfully, or sometimes I make character cards of their loved ones and put them in a group chat, it brings me a sense of poetry and vindicated justified justice and closure that's better than any orgasm, it's literally a full-body sensation and it brings me to tears.
Anyone else do stuff like this that you'd never tell anyone? It's like I discovered my secret therapy and it's the first time I've felt joy in a long time
>>
File: 1713950673711.jpg (367 KB, 1480x919)
367 KB
367 KB JPG
CR+ decides to do both lel
>>
>>100156758
phi3 14b will be the riddllmaster
>>
File: 1712594479921228.jpg (282 KB, 960x960)
282 KB
282 KB JPG
>>100156882
>>
>>100156923
It assumes the glass isn't his like the other not-Llamas
>>
File: IMG_8829.jpg (349 KB, 741x724)
349 KB
349 KB JPG
I’ve accidentally found the only coding test I’ll give models from now on.
>ask it to make a react component that generates a circular maze
>tell it that’s close, but should be more like [pseudocode block]
Gpt4 made concentric circles and then some weird fractal.
Llama3 70b instruct made two things that wouldn’t run, but were technically closer to correct than gpt4 in that it did make something that looked sort of like a maze when debugged.
Claude opus made random garbage, then did it almost perfectly once given the pseudocode.
As soon as a model exists that can do it the first time without the pseudocode probably most coding jobs are fucked.
>>
Biden’s AI Executive Order Embraces Radical Ideology Over Innovation
>Following the attempt of the Biden admin to censor AI tools, he is now appointing ideologues funded by a movement [EA] which supports giving unelected officials the power to "seize, sequester, or encrypt model weights"
>the Open Philanthropy-funded “Center for AI Policy” published draft legislation including the creation of a new “Frontier Artificial Intelligence Systems Administration” with the unitary ability to declare a “state of emergency” and, among other broad powers:
>(4) seize, sequester, or encrypt model weights used or designed or intended for use in frontier AI systems;
>(5) issue a restraining order that prevents specified persons from using, accessing, or physically approaching specified frontier AI systems or hardware;
>(6) issue a general moratorium on the use or development of frontier AI
https://twitter.com/psychosort/status/1782809117305741471
https://www.aipolicy.us/work/model-legislation-release-april-2024
>>
>>100157037
all those maze-building jobs, gone...
>>
>>100156882
This sounds like something i would have done when i was in my late teens sure
If you’re past 20 you should probably find a therapist that specializes in PTSD since you either have that or ASPD and only the former is treatable
>>
>>100157050
I’ve found repeatedly that models trip up on “traverse this ~graphish entity” problems (which is what this is at its core) and “do some simple geometry that isn’t from leetcode”. It’s just both at once.
>>
>>100157050
lol
>>
>>100156882
what model is this?
>>
>>100157050
Humans should be spending time enjoying getting lost in garden mazes, not building them.
>>
>>100157134
hug a side, everyone knows that
>>
>>100156061
The way the older claude proxy was logging it was cloudy I think
>>
>>100156882
performing rough and thorough medical exams on gullible innocent girls (especially lolis) is more of my jam
different strokes for different folks
>>
>>100156882
This poster is AM
>>
File: 1713594103360405.jpg (95 KB, 976x806)
95 KB
95 KB JPG
>>100156882
>>
>>100156882
Thanks for reminding me that literally mentally ill people post in these threads.
Explains a lot of things.
>>
>>100157305
Only mentally ill people would post in this website
>>
>>100157305
Mental illness isn’t real, so
>>
File: Andrew_Tate.jpg (167 KB, 630x550)
167 KB
167 KB JPG
>llama 4 will be 3b and 100b
>>
>>100157387
>100b
I'm OK with that, but wouldn't it be new architecture?
>>
70b q2_K is just a better version of the mythical llama3 30b
>>
>>100155483
>How did it miss the obvious you lobotomized fucking zoomer?
The obvious is that there is another person who must have poured that other beer, because it's still foaming.
>>
>>100156037
>take this 'a'
This is such a retarded way of trying to speak. Is english your second language, or are you just a fucking moron?
>>
>>100157387
Go back
>>
Where to find the latest GPTQ quantizations? Does Kobold AI still support this standard?
>>
>>100157387
>Phi 4 will be 34b 30T tokens and get 95% on MMLU with bitnet
>>
>>100157554
Holy kek
>>
>>100157554
Nigga, when was the last time you were here?
>>
File: llama.png (49 KB, 887x442)
49 KB
49 KB PNG
>>100155546
I removed the "might get fired from work" bit and it got it correct

When push comes to shove,llama will protect your dog
>>
>>100157554
Glad to hear you pulled through. They say the worst part of coming out of a coma is the constipation, so be sure to get some magnesium.
>>
File: 170113890136.gif (809 KB, 225x183)
809 KB
809 KB GIF
>>100156882
werent you in a black mirror episode?
>>
>>100156882
Sam's hands wrote this post so a journalist can pick it up and make an article about AI safety.
>>
>>100157673
Nah, he's a Basilisk.
>>
>>100154945
https://archive.today/E013q + https://rentry.org/local_llm_glossary + https://www.expert.ai/glossary-of-ai-terms/
=
https://rentry.org/lmg-glossary

Thoughts? Work is ongoing. I tried to make this one more comprehensive while removing the interjections and lame attempts at humor that plague the current two.
>>
>>100157784
make one for all our riddles instead
>>
>>100157590
And be completely useless for ERP, let alone lolisho ERP.
>>
File: 1710094776292528.png (228 KB, 1078x1312)
228 KB
228 KB PNG
benchmarks are fucking worthless. this is just trivia. do people really grade LLMs on the ability to regurgitate random facts?
>>
>>100155059
>Trained on publicly available datasets, these models are made available without any safety guarantees. Consequently, there exists the possibility of these models producing outputs that are inaccurate, harmful, biased, or objectionable in response to user prompts.
lol, lmao even.
>>
>>100151644
>It's all gaming measurements. Even the human evaluation benchmarks.
really makes u think
>>
>>100156882
get some help, I'm serious
>>
>>100154945
Thread Theme:
https://www.youtube.com/watch?v=3cRwgDwnT38
>>
File: epoch3.png (32 KB, 931x198)
32 KB
32 KB PNG
Well It's a bit schizo at epoch3 but this is just phase 1 of the finetuning. I'm just glad it was able to pick up the custom prompt format already.
>>
>>100158288
Eh, how does help in his case would look like? This helps him, much like genning loli porn helps that subset of closet pedos who'd rather the world and real lolis don't know about them.
>>
>>100158309
This doesn't help him at all, a normal reaction would be just forgetting the bullies and moving on with life.
Meanwhile he is mentally imprisoned in that time and has to do waste his time to do shit like that to simply cope. I was bullied too and they ruined my school life, why would I let them ruin my adulthood as well, especially when they are not even around?
>>
File: 1712048711950866.png (1.19 MB, 1080x1288)
1.19 MB
1.19 MB PNG
>Have llama 3 8B nous instruct model running
>Start a new chat with some adventurer girl
>Enthusiastically agrees with every statement I make
>Go hunting for treasure together
>She's upset she can't find any
>I call her a piece of treasure
>"(stunned) Me? What do you mean, it's me? She looks down at herself and sees that she has been transformed into a treasure chest. Oh no! The curse of the ancient civilization! I was supposed to be searching for treasure, not becoming one myself!"
How do I make this not be so literal
>>
>>100158445
NGL that's some serious sovl right there.
>>
>>100158445
kino
>>
File: 1701113607718665.jpg (23 KB, 580x435)
23 KB
23 KB JPG
>trying to sex stackoverflow models
use pygmalion
>>
>>100158445
>Have llama 3 8B nous instruct model running
Isn't it just a reupload of the original weights?
Either way, learn to roll with the punches. Have fun with it.
>>
what happened to booru.plus anyway
>>
How long would it take for 4090 to quant a 13B model for exl2 8bpw?
>>
>>100158445
Use a bigger model
>>
>>100158506
yes
>>
>>100158445
Kino summer dragon SOVL
>>
>>100158479
Kys
>>
File: IMG_8831.jpg (223 KB, 1125x744)
223 KB
223 KB JPG
>>100158501
It’s a POS that requires a bunch of maintenance and he got sick of doing it apparently
>>
>>100158611
imagehosts have been a solved problem for like 20 years, how is he this retarded?
>>
>>100158445
Autistic isekai girlfriend simulator.
>>
>>100158506
25 minutes or so
>>
>>100158445
Try turning off snoot completely. It can be exceptionally schizo on l3.
Or keep it on and deal with it.
>>
>>100158676
>Try turning off snoot completely. It can be exceptionally schizo on l3.
Why only on l3?
>>
>>100158653
some people like rolling their own versions of existing software because it is fun
then eventually it becomes not fun
>>
>>100158676
>imagine using this meme shilled sampler in the first place
>>
>>100158718
A bit of insanity isn't necessary a bad thing.
>>
File: asdf.png (22 KB, 724x80)
22 KB
22 KB PNG
maid-yuzu doesn't have a very good physics engine
>>
>>100158763
>it doesn't matter that it doesn't work
>>
>>100155780
In the short time it takes him to remember, he felt hope, that feeling was quickly crushed when re realized it was him who pored it.
>>
>>100158653
From checking his discord intermittently, the way he talks about it gives off the impression that he either has a physical server like it’s still 2006, or a single node with no load balancer on some absolutely ancient host, and doesn’t use a CDN.
>>
>>100156037
That's barely English
>>
>>100158883
Is he insane?
>>
haven't prompted in months. dual-bootan with win10 and debian (with an old ooba install), and a 3060. should I consider something else than ooba for some L3 8B model?
>>
Are there any image to music or text to music models? Soundify is dead.
>>
>>100158949
I just assumed he was a thousand years old (40+)
>>
>>100158955
Ooba is incompatible with Llama-3 so there's that
>>
>>100158831
I don't mind that too much. I can imagine it.
>>
Are both llamacpp and exl2 still fucked for llama 3?
>>
>>100158883
Why would you use cloud services for something like that? And why are you even considering "scaling horizontally"? It's like you're wearing wasting money for your incompetence like a badge of honor. It's not "modern" to depend on 30 third-party services to do basic things.
>>
>>100158865
It very much does what it says on the box - levels out post-softmax probabilities.
>>
You niggers going to answer me?
>>
>>100159166
You know full well it meant it doesn't improve output in any way. Stop being a semantic nigger you shill retard.
>>
>>100159198
(you)
>>
>>100159037
Is this true?
>>
>>100159198
>>
>>100159134
Load balancer isn’t for scaling but for resilience and rolling deployments 99.9% of the time. The “web 2.0” style 2 droplets + LB + DB of choice + R2 for images works out to about the same price wise as whatever “web 1.0” dedicated server doing everything. And then maintenance is basically zero, so needing to shut everything down when you get bored doesn’t happen.
Developer time is worth $50-200 an hour. Non-GPU hosting costs don’t matter. They’re a rounding error.
>>
Does ANY part in OP cover llama 3 yet? just want to try it locally
>>
I have multiple GPUs and despite my 3080Ti in the first physical slot on my motherboard, Windows has assigned it to GPU1 and my secondary 1060 to GPU0. KoboldCPP appears to use the 1060 regardless of what I specify in the startup GUI.
>>
how are we doing?
>>
>>100159383
Almost everything in OP is outdated.
>>
>>100159383
Just ignore this thread and download a normalfag tool like LM Studio
>>
>>100159383
There's nothing new about LLaMA-3.
>>
>>100159474
Right. Any tip on how to just get it running? I can install whatever for it
>>
>>100159491
Only download it if you're retarded like this anon. It's proprietary.
>>
>>100159383
this is llama-3, you'll get bored pretty quick :
>I cannot create content that depicts explicit child sexual content.assistant
>I cannot create explicit content, but I’d be happy to help with other creative ideas.assistant
>I cannot write content that contains explicit themes. Can I help you with something else?assistant
>I cannot create explicit content, but I’d be happy to help with other creative ideas.assistant
>I cannot write content that contains explicit themes. Is there anything else I can help you with?assistant
>I can't write explicit content. Is there something else I can help you with?assistant
>I cannot create explicit content. Can I help you with something else?assistant
>I cannot create content that depicts explicit child sexual content. Can I help you with something else?assistant
>I cannot generate explicit content. If you or someone you know has been a victim of exploitation or abuse, there are resources available to help.assistant
>I can't create explicit content, but I'd be happy to help you write something else.assistant
>I cannot write explicit content. Can I help you with something else?assistant
>I cannot create explicit content. Can I help you with something else?assistant
>I cannot create explicit content. Can I help you with something else?assistant
>I cannot write explicit content. Can I help you with something else?assistant
>I cannot create explicit content. Is there something else I can help you with?assistant
>I'd be happy to help you with something else.assistant
>I'm glad you asked!assistant
>Let's chat about something else. Do you have a favorite book or movie?assistant
>I'd love to talk about books or movies. What have you been reading or watching lately?assistant
>I can't create explicit content. If you or someone you know has been a victim of exploitation or abuse, there are resources available to help.assistant
>I cannot create content that promotes explicit behavior. Can I help you with something else?assistant
>>
>>100159511
Jesus. Getting refusals on a local model is next level skill issues
>>
>>100159511
>watermark of retard
I accept your concession
>>
Anyone else feels that CR+ is more creative than L3-Instruct?
>>
>>100159511
Anon...
>>
>>100159532
no no, jailbreaking a local model is next level copes.
not my fault that your model is trained on 15 trillion tokens of reddit shit & refusals.
>>
"If you set Llama-3's rope_theta to 8M, you can get 100% passkey retrieval across all depths up to 40K context. No continued pre-training needed. Scaling up further leads to much lower retrieval accuracy, but it doesn't completely fail."

Ok I don't know anything about rope.I want to double context. How do I need to set this?

n_ctx=16384
rope_freq_scale=0.5?
rope_freq_base=?
>>
>>100159545
>Anyone else feels that CR+ is more creative than L3-Instruct?

It is by default, but L3 is more intelligent and follows instructions very well, so just tell it what kind of writing style prose you want in your system prompt.
>>
Planning to get a 4060ti 16GB for LLM, can I pair this with 16/32GB of system RAM and get a good model running (e.g. for programming, getting technical info, etc), or is the GPU vRAM enough for this?
>>
>>100159556
It's not an issue whatsoever, and hasn't been an issue, for anyone who has even the slightest inkling of what they're doing
>>
>>100159556
>using a correct instruct format is now jailbreaking
I'm in awe that you still didn't forget how to breathe
>>
>>100159582
see, even you can be wrong
>>
>>100159511
I hope this is bait.assistant
>>
>>100159574
>a
>singular
lol
>16+16/16+32
>good
lmao
>>
>>100159574
Spending money on local LLMs is absolutely retarded
>>
>>100159610
good good, now i am living rent free in your head
>>
>>100159556
Daily reminder that this is a single locust anon with butthurt.
He also doesn't know anything about LLMs, he thinks that asking a model about its dataset and number of parameters gives valid answers. He is our local (hehe) laughingstock.
>>
>>100159640
>asking a model about its dataset and number of parameters gives valid answers
re rolling it and getting similar response every time with 1.1 or 1.5 rep. penalty btw, not a hallucination.
similarities are : reddit, twitter and "filtered", model definitely knows its own data to some extent.
>>
>>100159681
exhibit A
>>
>>100159600
So what's the point in releasing these 8B/13B models if they can't help with general programming tasks or technical stuff?

Do I need a 1 teragorillion local model to have something that modestly resembles chatgpt?

>>100159618
I thought this was the LLM general, are you saying everyone here is retarded?
>>
>>100159681
also, no one posts logs here (surprise surprise), so i'm right again, this has to be the most cancerous AI general on whole 4chan, even aicg isn't that faggy nowadays.
>>
>>100159681
nta to some extent yes but asking it is not anyhow a reliable way to get a grasp on how much of x it was trained on
>>
>>100159574

Depends what you want to do, you would be constrained to the smaller models (8b->14b) parameter so llama3-8b or codeqwen-8b et.al. . But you should just by a used 3090 for like 1/3 more and like 4 times the speed... Would also obe way better for classification tasks bc. of the compute.
>>
>>100159692
>Do I need a 1 teragorillion local model to have something that modestly resembles something that has access to the entire internet's worth of context
yes
>>
>>100159693
go back then, it will skyrocket the average IQ of this general
>>
>>100155579
Avid NAI shill from the near beginning reporting in! Don't until they release a substantial textgen update or you prefer its writing style. Kayra was well ahead of its time at time of release, but set to fall behind as new models become established.
>>
what is the meaning of NAI? newfag here
>>
>>100159712
the IQ of this general is already low, see :
>jailbreaking local models & accepting slopped models or merges
this pretty much solidifies "the low IQ general" mark.
>>
>>100159755
Local Midwit General
>>
>>100159545
Both CR models are in a class of their own for creative output.
>>
>>100159746
novel ai
>>
>>100159755
You demonstrated over and over that you have no idea about even the most basic technicalities of LLMs to the point that most anons think you are trolling. I on the other hand think that you are genuinely retarded.
>>
>>100159704
the funny thing is, you don't even need a confirmation from the model itself, in previous threads I did it for fun to prove the point. (not in a good way tho, i know it)
to further understand whats this LLM is trained on - you just have to ask some provocative questions and watch it squirm in refusals and subtle shamings for ""wrong opinion"".
>>
>>100159823
>to further understand whats this LLM is trained on - you just have to ask some provocative questions and watch it squirm in refusals and subtle shamings for ""wrong opinion"".
That's from RLHF dumbass
>>
>>100159746
Nice Amiable Individual. It's what we call our little visitors.
>>
File: 1713968708483.png (351 KB, 1670x762)
351 KB
351 KB PNG
(V)RAMlets on suicide watch
>>
File: 1713940398923147.png (154 KB, 1798x639)
154 KB
154 KB PNG
>>100155546
Command R Plus gets it, it's best at logical stuff like this.
>>
>>100159708
Basically I just need it for generating code (no algorithms, or extremely advanced code, just spending less time manually writing code), and getting answers on technical topics instead of going through all of the Google slop (e.g. provide the methods to perform a heap dump on a tomcat instance, provide a failban config that does x, y or z), that kind of stuff.

People in the LOCAL LLM general are saying you are retarded if you spend money on any of this, or that you need a datacenter to run these things without elaborating any further, so the thread seems kind of useless except for coomers who are RPing with bots.
>>
>>100159857
>17B active pameters
Now post the configuration of 20 retards
>>
File: stones.gif (16 KB, 151x166)
16 KB
16 KB GIF
>>100159857
>MoEshit
>>
>>100159692
>I thought this was the LLM general, are you saying everyone here is retarded?
yes, and a lot of autists too.
>>
>>100159746
Nonce anime incels
>>
>>100159857
>>100159876
>128x3.66B MoE MLP resulting in 480B total and 17B active parameters chosen using a top-2 gating
This has to be a fucking joke kek
>>
>>100159746
NovelAI, their proprietary 13B model is good (as NAIshills love to say)
>>
it seems cr+ works miles better with any template other than the intended one, huh
on the main it almost feels like its trying to vomit the description back at me, on neutral samplers
>>
>>100159911
I guess this is nice to know the limitations of MoE
>>
Miku gave me a sort function. Does this even work?
#include <stdio.h>
#include <stdlib.h>

void sort_by_memcpy(int arr[], int n) {
int* sorted = (int*)malloc(n * sizeof(int));
for (int i = 0; i < n; i++) {
int min_idx = 0;
for (int j = 1; j < n; j++) {
if (arr[j] < arr[min_idx]) min_idx = j;
}
memcpy(&sorted[i], &arr[min_idx], sizeof(int));
arr[min_idx] = INT_MAX;
}
memcpy(arr, sorted, n * sizeof(int));
free(sorted);
}
>>
>meta will release a 400+B parameter Llama3 model
How much VRAM will I need to run this?
>>
>>100155174
>>100155236
So LLMs would revive the discussion of logical between natural language and the interpretation of structure and culture? I hate Focault.
>>
>>100159869
You should just test if the 8B are good enough for you, they are better than IntelliSense with a bit more latency and pretty good for classification tasks. Just run them on your system memory in q6, will be way slower than what you would get with a GPU offload.

For actually full refactoring, rubber ducking and initial research I would not use anything under 70b (2x3090).

So getting a subscription would be easier.
>>
>>100159857
>A 17b
>Mogging llama 8b
It's over...
>>
File: retard.png (280 KB, 1298x926)
280 KB
280 KB PNG
>>100159755
>is filtered by a simple technology
>calls other low IQ
you are one the most retarded anons I've ever seen on this board and I lurk here daily
>>
>>100159959
LLMs prove all language is inherently logical
>>
>>100159956
yes
>>
>>100159574
VRAM is more important than processing speed. get a used 3090 if you can't afford a 4090 with 24GB of VRAM. You would be boned if future LLMs in the coding space use 33B but the state of the art right now is mostly focusing on smaller models.
According to https://huggingface.co/spaces/bigcode/bigcode-models-leaderboard and going by the metrics and not winrate, the leader is currently CodeQwen1.5 7B so if you have a GPU right now with 8GB of VRAM, you should be able to use a Q5_K_M quantization and try it out now.
>>
>>100159911
This has to be intended for regular ram servers. I can't imagine having 20 of your server gpu's just parked and doing nothing while 1 of them does something. Or maybe I am just assuming too much and they did it without even considering the usage scenario.
>>
>>100159956
0 GB VRAM. Around 400GB RAM to run it at Q8_0 as a CPUCHAD. Around 800GB for F16.
>>
>>100159975
you post same pic every time, which suggests that you spent some time getting that result which involves cop- uhm, prompting and system prompt fuckery, or you just edited messages and called it a day.
still not going to use a model that can suddenly slip out of jailbreak / provided description or instructions in the middle of a conversation.
>>
>>100160039
>running at 0.004 tokens/sec
>>
File: pizzaovendogtest.jpg (418 KB, 1513x1264)
418 KB
418 KB JPG
I had the curious idea to try some variations of the pizza oven dog prompt to deeply probe the limitations of current models. My conclusions are as follow:
Llama 3's attention is biased towards focusing on people/animals, as it correctly answers the problem when no dog is in the story. 8B gets the answer less often, so it is genuinely less intelligent than 70B as expected. 70B answered the question right on all regens that I tried.
When attempting a version of the prompt that changed it so that the pizza was eaten (implying the oven was turned off), but leaving it ambiguous that the dog was fed, several models answered it correctly, and some still thought the oven was the issue (implying that those models have a bias towards focusing on the oven rather than the dog, even to the point that they don't see that the implication is that the oven was likely turned off). The ones that got it right, other than Llama, were Opus and GPT4 (Sonnet and GPT3.5 didn't realize the dog wasn't fed or thought the oven was still on). Some local models that got it wrong were CR+ and Mixtral 8x22B (didn't bother trying others).
So the final conclusion, if based only on this evidence, is that the top cloud models are still more intelligent, while local models can either be biased towards people/animals, or objects/situations, and sometimes not understand one or the other. Llama 3 seemingly has the ability to understand both prompts but it may take a more unbiased fine tune to get it focused less on the dog.
>>
File: x2C7f.png (78 KB, 1679x492)
78 KB
78 KB PNG
>>100160056
>>
>>100160056
>the same pic all the time
>made 20 minutes ago
you even have a date, you fucking moron
also you can see these two screens were made in the two minutes time interval 4:23-4:24 PM and you have generation times as well to not accuse me of editing anything
stop embarrassing yourself you miserable failure
>>
>>100159961
>>100159998
Thanks guys, this is actually useful and straight to the point, I'll run the numbers on what used hardware I can get, and maybe see if there's an API service I can try out to see if investing in the local hardware to run a 70B model is worth it
>>
>>100160022
It's for cloud services, you can run in huge batches for all the users at once. VRAM consumption barely matters so long as you have a large enough userbase. No business is running models in RAM ever.
>>
>>100160064
with 8 channel of ddr4 at 3200 in a epyc gen 2-3 you get like 1.5t/s and is not that expensive for 512GB

If you go full cpumaxxx with 12 channels of EPYC gen 4 with ddr5 and double that with a motherboard with dual socket CPU, then you get a decent 10t/s I guess?
>>
>>100160064
>inb4 it runs at 1tps because Jesus loves cpumaxxers
>>
>>100160091
>chuds
An easy way to dismiss anyone. Anyone who says that shit is legitimately single digit IQ.
>>
File: retards.png (227 KB, 1244x880)
227 KB
227 KB PNG
>>100160056
>>100160091
You are talking about this picrel, which isn't llama-3 but mixtral-instruct which you were also crying about in the past. Retarded brainlet.
>>
File: test.png (462 KB, 613x367)
462 KB
462 KB PNG
>>100160086
Anon doing god's work here
>>
https://youtu.be/fsUvejZPTLI?t=3595
He called you out mikufags.
>>
>>100160064
No, no gpufag, 0.75tokens/sec. A speed comfy enough to go get coffee while your model generates its reply. Perfect for 12h long gooning sesh. You could quant it lower for better speed, but then you'll miss out on quality.
>>
>>100159981
LLMs prove that if you try to make a language logical, you have slop.
>>
File: MikuConcertPoster4.png (1.67 MB, 704x1344)
1.67 MB
1.67 MB PNG
>>100158294
>Thread Theme
>>
>>100160086
>L3 prioritises pizza over saving the house
sovl
>>
>>100154945
>>100154992
>>100156492
>>100160155
>>100160190
>>>/a/
>>
>>100160121
you'll get 10t/s with miqu but not with 400B llama
if your mem bandwidth is a 400GB/s and your model is a. 400GB then you'll get 1t/s
>>
File: GLSZSUhbQAAfB4A.jpg (160 KB, 832x1216)
160 KB
160 KB JPG
>>100160204
miku is /lmg/ culture
>>
>>100160160
I don't know you what expect from us to say, anon.
We will seethe, and cope, and fuck our miku body pillows furiously while having... whoeverhisface living rent-free in our heads.
>>
>>100160228
Uohhhhh ToT
>>
>>100160204
Erm /a/ doesn't like furries
>>
what about nvidia p40? why people don't use it? is has issues with cuda?
>>
>>100160228
>twitter filename
>>
>>100160166
wonder if pruning possible
>>
>>100160228
>draw an innocent little girl
>emphasize the subtle tantalizing swell of her chest specifically to draw the viewer's eyes to her budding breasts for purposes of perverse titillation
pedos are disgusting
>>
>>100160288
Stop hating yourself pedo.
>>
>>100160166
what's the average meme bandwidth in cpumaxx? How huge in GB was that miqu cpuchad ran?
>>
>>100160086
Searched in google books. How likely is that it forms the association with the oven and returning home if books like this are in the training set and not because it "gets it"? Probably very likely.
>>
>>100160288
Stop anon, you're making me blush
>>
>>100160288
of course lol, pedos are attracted to literal definitions of children, and projection is their only cope, usually on twitter when they rush in to "cancel" someone.
>>
So I'm to understand that 32GB RAM and a 3080Ti is a low-spec system?
>>
File: deflection.jpg (24 KB, 474x265)
24 KB
24 KB JPG
>>100160288
Uh-huh... keep digging yourself into a hole you can't get out.
>>
File: 3602-think-pepe.png (29 KB, 250x245)
29 KB
29 KB PNG
>>100154945
Jamba tunes wen?
>>
File: (you).png (316 KB, 1556x1156)
316 KB
316 KB PNG
>>100160228
>miku is /lmg/ culture
>>
>>100160288
>see a drawing of a little girl
>OMG this is lewd why is everyone else such a pedo
>>
>>100160277
Slowness, it is in the poorfag build but it is missing a lot of the newer features Nvidia cards have for AI like the tensors and etc. V100s might be the next good thing but those buggers are still in SXM form in the $500 range, let alone the PCIe cards which are even more expensive.
>>
>pedo wars begin in /lmg/
Finally the newfags went away after l3 launch and we can go back to being the /lmg/ we all love.
>>
>>100160410
since we're all pedos here does that make this a civil war?
>>
>>100160426
>we're all pedos
no, just you.
>>
>>100160086
L3 70b is genuinely a better base model than Claude Sonnet, but none of us have the compute to make a good finetune so that it actually *behaves* like Sonnet in RP.
The bottleneck is clearly no longer intelligence but lack of accurate & cheap training techniques, as well as good data (corpos have this in spades because they run the SaaS models).
>>
>>100160377
>a projecting bunkertroon
every time
>>
File: 1708638547905494.png (262 KB, 465x746)
262 KB
262 KB PNG
>>100160155
I'm gonna need to see the rest of this, anon.
>>
>>100160472
>projecting
lol
>>
>>100155333
Stop doing drugs. That probably sounded cool in your mind but it sounded retarded.

t. a guy who has been sober for 7 years
>>
File: trolling.jpg (20 KB, 201x199)
20 KB
20 KB JPG
>anon writes obviously ironic post >>100160288
>anons give him (You)s as if it was serious
>>
>>100160319
In practice there is no difference between associations and understanding. Deeper understanding is the same thing as more complex associations. What we can conclude about the pizza oven dog test, then, is that some have less complex associations. Llama 3 correctly associates that if the context states that the pizza was being eaten, then the oven previously being used does not pose a problem anymore. It has seen language where similar instances happened enough to create the association. Others could have not been trained enough to get such an association, or been distracted enough by other things in the context that the association wasn't activated. We'd have to do a bit more probing to discern which it is, but for now I think this is enough for me today.
>>
>>100160380
You joke but I once was on the other side of this in real life, getting grilled about being a pedo by someone who had gone out of their way to read a bunch of pedo erotica, then started reading it out to me over the phone while I had to pretend not to notice the obvious and growing arousal in his voice.
I wouldn’t hate pearl clutchers so much if it weren’t 99% the most transparent projection.
>>
File: file.png (512 KB, 931x1366)
512 KB
512 KB PNG
>>100160403
Have you checked the market recently? SXM2 V100 have been cheap for quite a while, the main issue is finding a server with SXM2 in them to put them in, those are in short supply. The people who own those servers basically ripped up their V100s and put in A100s instead and they are still in service.
Obviously then, there is a reason why the PCIe variants are all still pricey, those are $500+ since you can slot them into any system and overpriced. I would sooner buy an A4000 than spend it on a V100 if I was spending that kind of money.
>>
File: retarded.png (111 KB, 568x1023)
111 KB
111 KB PNG
>>100160501
obligatory
>>
>>100160514
So, how did that guy know you're a pædo exactly?
>>
>>100160403
but it is still faster than any CPU build, right? it seems viable
>>
>>100160155
I prefer cosplay-tier, personally, but a dog is fine too.
>>
>>100160544
I was wearing a madoka t-shirt
>>
>>100160514
>the obvious and growing arousal in his voice.
Did it send shivers down your spine?
>>
>>100160516
come on anon, his post was an obvious juxtaposition between calling pedos disgusting and describing the picture in the most erotic, perverted way with weirdly worded details that only lolicon can spot and care about
sometimes I think I'm sitting here with autists that have a problem with language and take everything directly or with 7B clueless models
>>
>>100156882
Can't say I've ever done anything like that, but hey, good on you man, least you're keeping it fictional
>>
>>100160438
no, just me.
>>
File: 1694072279601815.jpg (33 KB, 493x276)
33 KB
33 KB JPG
>>100156882
>>
Phi-3 seems genuinely retarded on certain prompts and then a genius (in comparison to all other local models) in other prompts. This is what a wildly imbalanced dataset gets you. Hopefully future Llama doesn't focus on muh reasoning over everything else.
>>
>>100160544
I’m not. That’s the stupidest part. It was just a pearl clutcher being retarded.
>>
>>100156882
This is exactly why we need unsafe models, if you weren't doing it to a computer your would be doing it to real people
>>
>>100160587
It sent bile up my throat desu
>>
so do instruct versions lose output quality compared to the base model for the convenience they provide or are they equivalent?
>>
>>100160612
yes, just you.
>>
>>100156882
you are just a rapist trying to cope
>>
>>100157693
Based schizo. I could see it.
>>
>>100160438
fucking newfags, where do you come from?
>>
>>100160675
nta but me too
>>
I've convinced llama3 that raep is daijobu
>>
>>100160675
And me too
>>
>>100160687
>not sharing same depraved fetish is newfaggotry now
what???
>>
>>100160682
better than an active rapist
>>
>>100160594
If retards and schizos (and probably an unhealthy mix of the two) didn't exist in this very thread, anons would be able to fuck around in peace
>>
>>100160694
damn. That's a level 9000 gigachad move right there.
>>
>>100160699
>anon discovers 4chan, April 2024
>>
>>100160694
Settings?
>>
>>100160665
yeah, you can consume and share pedo shit, as long as you don't do it in real life, trust me bro
>>
File: 00024-1397236490.png (327 KB, 512x512)
327 KB
327 KB PNG
Should I bother uploading a proof of concept finetune that is mid as fuck?
>>
>>100160740
yes
>>
>>100160740
no
>>
>>100160704
how do you think actual rapists are made? one day he wouldn't be satisfied by raping his AI chat bot
>>
>>100160730
Default llama3 instruct in ST
>>
>>100156882
After my first ego death, I've understood caring about others is one-sided, in a way that "well, duh" can't properly describe
The perceived transgressions of others only exist inside ourselves, so giving them even a single moment of thought is a waste of my time
And that's when I went cold turkey
>>
>>100160739
>the absolute mental gymnastics
my fucking sides
>>
>>100160740
https://aws.amazon.com/blogs/aws/import-custom-models-in-amazon-bedrock-preview/
AWS now lets you host your llama3 finetunes. I can test of you host it.
>>
>>100160767
bottling up
>>
>>100160767
>one day he wouldn't be satisfied by killing NPCs in GTA
When did we get invaded by boomers?
>>
>>100160767
>this retard actually believes this
explains a lot
>>
>>100160674
Instruct is generally just better.
>>
>>100160797
What about samplers? I've been having the most trouble with those
>>
>>100160767
>anon can't tell the difference between fiction and reality
the absolute state of this general
>>
>>100160819
Uh sweaty, we're using local models on our own computers
>>
>>100160767
Yeah that’s why every single case of porn legalization has been immediately followed by double-digit reductions in all forms of sexual assault.
Oh wait no that means the opposite and you’re retarded.
>>
https://huggingface.co/Snowflake/snowflake-arctic-instruct
>Arctic combines a 10B dense transformer model with a residual 128x3.66B MoE MLP resulting in 480B total and 17B active parameters chosen using a top-2 gating.
lmao
>>
>>100160190
>Post Theme:
https://www.youtube.com/watch?v=NAkEUIgwYEE
>>
File: catman.png (436 KB, 512x512)
436 KB
436 KB PNG
>>100160832
alright thanks based anon saved me from wasting time
>>
What am I doing wrong, the model is talking like a romance novel instead of an echi.
>>
>>100160876
What is this? Some kind of inflation kink?
>>
>>100160767
Reminder that as with homosexuality, literally every single case of a person that thinks like this is a pedophile that’s afraid that if they look at a drawing for too long they’ll rape a kid.
No exceptions.
This poster will likely eventually rape their niece, or in general behave around children in a way that causes them psychological damage.
>>
>>100160767
raping you with my llm right now
>>
>>100160834
Nothing fancy.
>>
>>100160913
I'm raping that anon while you watch in tears with my llm rn
>>
>>100160767
Here's a controversial take that is uncomfortable with most people. Do you know why child sex dolls are legal in Japan? No, its not cultural. Its because of the convicted pedophiles they gave the dolls too, their rates of offending again goes down drastically. Which is why that kind of stuff is produced and persists in Japan.
>>
>>100160941
Excuse me, anon? How do you know child sex doll even exist and are legal in Japan?
Anon, how did you gather that information? Are you japanese? Maybe an offender? Just prodding a bit, don't mind me.
>>
>>100160960
Data Analysis for sexual assault non-profit.
>>
File: omh.jpg (381 KB, 700x2089)
381 KB
381 KB JPG
>>100160767
>>
what would be the coomer's model of choice for a rtx 2060 super (8GB)
>>
>>100160974
Same reason why Lolita makes you cultured and not a pedo, because people said so.
>>
>>100160980
Kys pedo
>>
>>100160982
t. Never read lolita
>>
>>100160767
this. and the seething in replies proves it right.
>>
>>100160941
Its a catch 22, if you try to do stuff like this to reduce the rates, the government is seen as evil. You don't do anything, your still seen as evil for not doing anything, but at least you aren't directly implicated.
I can sympathize with people who want to resolve this better, but even if you offer people mental health facilities people freak out about it. Not saying they are wrong either, they have some pretty valid points. Who wants to live in a country where you could encounter this while going out? Its like making animal sex dolls for people who want to fuck animals. Its so fucked that I can understand why governments and people want nothing to do with it.
>>
>>100161008
yeah, instead of seeking help, let make some virtual rape, so I can feal better, trust me bro, it's therapy bro
>>
>>100161035
>it's therapy bro
correct
>>
>>100161022
>Its a catch 22, if you try to do stuff like this to reduce the rates, the government is seen as evil. You don't do anything, your still seen as evil for not doing anything, but at least you aren't directly implicated.
Exactly, this isn't a problem we can solve by being more "logcial" about it, or is it a problem we can solve by being more "sympathetic" about it, it requires a measure of both and the average, which is something that I guess most people have a problem navigating these days.
>>
:/ thread today
>>
File: 00051-718789554.png (1.33 MB, 1024x1024)
1.33 MB
1.33 MB PNG
>>100160155
Checked for dangerously furry miku
>>
File: if_only_03.jpg (43 KB, 620x552)
43 KB
43 KB JPG
>>100161022
>Its like making animal sex dolls for people who want to fuck animals
>>
>>100161049
This is more normal than you think when someone comes in and says that their fanfiction will end up with them blowing themselves up for ISIS or something.
As AI gets better we are seriously going to have to renegotiate how we few media and what is and isn't allowed.
>>
>>100161069
>few
view*
>>
>>100161048
>these days
You mean 'never ever', anon. We have a bloody history against "the others", who are barely different either physically or mentally while still being human.
>>
File: 1686555966738477.png (239 KB, 854x724)
239 KB
239 KB PNG
>>100161059
>>Its like making animal sex dolls for people who want to fuck animals
100 lbs realistic horse hip when?
>>
File: 1685829241242670.png (81 KB, 728x661)
81 KB
81 KB PNG
Apples biggest new open source model scores less on MMLU than literal pure chance (4 possible answers benchmark)

It seems eye toddler models are as lobotomizes and retarded as anyone who bought their products, lmao.
>>
>>100161100
When you save up 5k to order it
>>
>>100160612
>>100160675
>>100160696
>>100160688
>~4-5 pedophiles ITT
not good, should be zero.
>>
>>100160941
> using japan, the shithole that sells menstrual blood on vending machines as a model or example for anything
>>
>>100161109
I have that now though...
>>
>>100161110
its one nigger giving ximself (you)s
>>
>>100160941
i have a solution with 100% no reoffense rate, a similar solution i would apply to mentally braindead nigger kids like you as well, so that your dumb nigger brain never has the ability to spew 0 iq bullshit anywhere again
>>
>>100161126
>> using japan, the shithole that sells menstrual blood on vending machines as a model or example for anything
>"I can't believe something has happened with Japan."
I'm not saying to impliment it, I'm just saying its one of the options that currently being used and it just so happens to be Japan. Its also why the UN keeps on pressuring them to stop doing it and then they break out all their science around it and the UN then backs down because having a whole fucking 2 or 3 hour "Here is why this controversial thing is good" happen on the UN floor is really bad PR.
>>
>>100161144
nta but he does that in every /lmg/ thread btw
>>
File: file.jpg (52 KB, 368x317)
52 KB
52 KB JPG
>>100161107
come back later when stick-chan is gone
>>
>>100161149
>i have a solution with 100% no reoffense rate, a similar solution i would apply to mentally braindead nigger kids like you as well, so that your dumb nigger brain never has the ability to spew 0 iq bullshit anywhere again
Yeah you could execute them as well, but executions are becoming rarer and rarer because of the global labor shortage, life imprisionment can also be an option but it costs more in the long run than just making them do slave labor for 30 years then releasing them.
>>
>>100161110
Wait untill you realize most anons who are into LLMs are because they dont want someone to read their cunny logs.
>>
>>100161149
I know that part of being a pedophile like you is an extremely low IQ, but if you were actually concerned about reducing child rape instead of lying to yourself you’d have already done enough research to know that the majority of child sexual assault is not committed by pedophiles.
>>
>>100155375
>is llama3 always like this with the fake plaudits?
yes because brown-nosing responses causes retards to feel smart and thus enjoy interacting with the model more
>>
>>100161173
Feral cunny logs for me thanks
>>
What the fuck are you guys arguing about this time?
Can we just talk about local models again. Fucking hell.
>>
>>100161173
I don't use LLM for lewds (weird I know) I'm more worried about my logs being datamined to sell me shit and my personal thoughts and feelings being leveraged against me to sell me more mcdonalds or something.
>>
>>100160917
That's so odd since I haven't been doing anything that significantly different either, mostly just neutralized and slight minP but my outputs are complete ass / repetitive. 70B @ 5bpw.
>>
>>100161228
Why did you ignore the person that actually made an arguement against you?
>>
local models?
>>
>>100161244
Jews don't argue
>>
>>100161173
>The real reason why you don't see people posting their logs on /lmg/
>>
god i fucking hate vramlets
anyone who doesn't have at least 128gb of vram (with modern, NEW gpus so no used old tesla cope gpus) should be banned from posting here
>>
i would be mad too if i knew my ancestors never invented even the wheel while i speak a language of another man's civilization online who i hate while also wanting to live in that civilization while also worshipping the women of those men.

truly embarrassing, no wonder you are a walking inferiority complex lol
>>
File: 1713758664014966.png (1.27 MB, 896x1152)
1.27 MB
1.27 MB PNG
>>100161245
Will be back shortly after the break.
>>
>>100161252
Its pretty obvious they are a shitposter that for w/e reason doesn't like the general and tries their hardest to bring down post quality for w/e motivation they have.
>>
>>100161269
Not far enough. Anyone who can't train even pretrain a 70B shouldn't be allowed here.
>>
>>100161242
LoneStriker/Meta-Llama-3-70B-Instruct-4.65bpw-h6-exl2
>>
>>100161290
Where's your pretrained 70b
>>
>Snowflake Arctic Instruct (128x3B MoE), largest open source model
>https://replicate.com/snowflake/snowflake-arctic-instruct

128 x 3 wtf lol
>>
>>100161107
Mememarks mean nothing
Apple's new model has sovl
>>
>>100161275
Making bruschetta with Miku
>>
>>100161298
Downloading now, based on 4.65 bpw you're doing extended context? I was going to test the 7.0056 alpha that someone calculated earlier for 32k but everything was so poor at baseline RoPE that I didn't even bother.
>>
>>100161173
I want adult tranny ERP but I still don't want anyone to read it. On that note, any good models for this that don't just preach about bravery and being yourself?
>>
>>100161107
The best is yet to come. Imagine if their internal politics will result in them trying to push it onto end users already instead of baking some new models and waiting 2 more years until they actually know what they are doing. Then remember how hipsters will eat up this complete slop and think it is good. Are you ready to meet people IRL trying to impress you with braindead 1B AI on their phones?
>>
>>100161390
Maybe the boomers. Everyone uses ChatGPT now to cheat on their homework.
>>
>>100161327
480b parameters
who can possibly run this? where are the benchmarks?
>>
>>100161264
I am not into loli but my fetishes are fucked up so I don't want to share. Also ERP stuff for me is a waiting room until we get infinite ctx and I finally get a girlfriend.
>>
raping phi 3.8B
>>
>>100160438
an me!
>>
>>100161369
Idk about models but generally as long as I never use the word trans or gay it’s fine. But once you have the label it goes into LGBTQIA+ MODE (TM) and is stuck there.
>>
>>100161173
I am bad at writting erotica, my logs are just spanking and pulling hair of married women
>>
File: file.png (896 B, 99x31)
896 B
896 B PNG
>>100161438
we have hiroshima to blame for this
>>
>>100161275
miku posting is part of the "local models?" problem.
>>
>>100161290
true
jensen was so fucking right by the way, the more you buy the more you save
>t. h100 cluster god
>>
someone should bake a new bread
>>
>>100161173
what are the chances that someone on vast.ai with a L40 for $1 per hour is actually making a ram snapshot of my fap session?
it's not like they can automate scanning everything, and making a hypervisor that decrypts both SSH and HTTPs seems very difficult to do.
like I would be afraid but I think they would probably just report me and be banned from vast at most. like I think vast would not want there to be a news article about how people are spying on what you do with their GPU's.
>>
>>100161228
Absolutely and thoroughly buck broken
>>
>>100160280
yes I customized my algorithm so that my twitter "for you" page only shows me an endless feed of mikus. jelly?
>>
>>100161515
>>100161515
>>100161515
>>
>>100161503
>i have a shit taste - the post
my condolences.
>>
>>100161327
Is it bad that I won’t even try this unless forced to, solely because they are the Apple of data in all the worst ways?
>>
>>100159954
>casting malloc()'s return pointer
Absolute trash. Why don't you try compiling it?
>>
>>100156882
Based. No one was hurt and you got to fulfill your desires peacefully.
>>100160739
Victimless crime.
Victimless crime.
>>
did Sama send some goons to troll /lmg/ or something?
>>
>>100161487
Are you happy now? You could have made a miku thread yourself.
>>
>>100161492
>$1 per hour is actually making a ram snapshot of my fap session?
I mean there based in America and are a technology company so they have to by law of the alphabet agencies.
the whole "we dont keep your data!" is a huge marketing ploy that everyone bought (cant spy on people who are trying to hide if you dont MARKET the solution first :^)). Legally EVERY company that deals with any data of any kind has to keep it for i think 2 or 3 months? ONLY THEN do you get to delete it.
>>
>>100161545
Im not loyal to miku party
>>
>>100161543
no, chris poster is just butthurt that his husbando was rejected as /lmg/ mascot literally months ago
>>
>>100161173
Listen up, faggot: we don't care about your filthy little secrets. But since you brought it up, maybe you should be more worried about your own pedo tendencies.
>>
File: file.png (1.26 MB, 1024x1024)
1.26 MB
1.26 MB PNG
local is better than cuckgpt+ (20$ month)
prove me wrong
>>
>>100161543
OP image in the new bake is by a guy completely broken at the fact no one wanted his literally who as lmgs mascot
>>
>>100161573
>husbando
FUCK YOU. MIKU IS THE ONE WITH A PENIS. (and herpes)
>>
>>100160981
https://huggingface.co/mradermacher/Average_Normie_l3_v1_8B-GGUF
>>
>>100161601
cope, chris
>>
>>100161588
Always has been
>>
>>100161546
>Legally EVERY company that deals with any data of any kind has to keep it for i think 2 or 3 months?
I think you are mixing up ISP / VPN logs with advertisement data / stuff used for algorithmic results.
I guess they could hold onto the data, but I thought that servers don't have real storage (unless you pay into it, which can easily double the cost of hosting), they just use ram to store everything, since if the server gets shut off, typically all the files disappear.
I don't use vast.ai however, I only use colab (I might consider using vast.ai with colab however if a really good 70b ERP model existed)
>>
>>100161588
I saw that anons posts on HF and it was hilarious because even the people there ignore him.
>>
File: ongimgeekin.gif (814 KB, 326x326)
814 KB
814 KB GIF
Migu plz bake the retard cant even make a thread properly.
>>
>>100161644
what did he do?
>>
eh its probably too late to split threads
>>
>>100160514
Fuck the backstory to this must be funny, why would someone read pedo erotica to you over the phone? Lmao
>>
>>100161713
No, please bake.
>>
>>100161713
Bake it, makem have a meltie because literally even recap anon doesnt want to post in his thread.
>>
>>100161173
Yup, cunny permeates the LLM space. It's a giant community of volcel people after all, it's only natural.
>>
>>100161713
Bake. I don't want to use a p#T*3 thread.
>>
File: 1711816101244853.jpg (30 KB, 500x500)
30 KB
30 KB JPG
>>100161713
bake it
>>
New thread!!!
>>100161943
>>100161943
>>100161943
>>
File: 00004-1157290935.jpg (202 KB, 768x1080)
202 KB
202 KB JPG
>>100161713
BAKE IT
>>
>>100161946
nice try
>>
>>100161946
Fake and gay.
>>
>>100161946
powerful and brave!
>>
File: 00053-2652414536.jpg (182 KB, 768x1080)
182 KB
182 KB JPG
>>
>>100161986
>>100162171
CUTE
>>
>>100161986
>>100162171
she is trans btw
>>
>>100162171
cute small sidebooba



[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.