But... I only said hi.

[-]

-p-e-w-@reddit

FYI, Tiefighter is massively outdated. You can get smaller models that blow it out of the water. Try Stheno, or any NeMo finetune, or even vanilla NeMo. You’ll be amazed by how much more your machine can do.

Reply

[-]

dagerdev@reddit (OP)

Thanks for the info! I downloaded a bunch of models that people recommended in another thread and some were disappointing.

Reply

[-]

-p-e-w-@reddit

The Llama 2 architecture is ancient by now. They don’t use modern techniques like GQA, their tokenizer is crap, and pretraining was a fraction of its successors. The better finetunes like Tiefighter squeezed all they could out of that stone, but ultimately they all hit its inherent limitations. If you have 12 GB of VRAM, I recommend you run the new Mistral Small at IQ3_M quant. It’s close to GPT-4 in quality, and so far ahead of Llama 2 it’s hard to describe.

Reply

[-]

xor_2@reddit

Most modern LLMs are better than llama2. It was nice for the time being one of few open source models and all but these times these old models are pretty bad. Heck, even original GPT3 (as in very first models available in chat gpt site) is pretty pathetic compared to today's small models.

Reply

[-]

TheRealGentlefox@reddit

Wait, the new small fits at IQ3? How? Nemo just barely squeezes in at Q4.

Reply

[-]

TheRealGentlefox@reddit

As others have said, the meta is 100% Nemo right now. Unslop finetune is good, but it's hard to go wrong.

Reply

[-]

AppearanceHeavy6724@reddit

Finetunes generally suck; not a single has "blown me away", no matter what the redditors promised. But Nemo is good.

Reply

[-]

Such_Advantage_6949@reddit

Just stick to the official model from the provider. Most of the fine tune suck

Reply

[-]

ArsNeph@reddit

I'd recommend Mistral Nemo 12B instruct for general tasks and Mag Mell 12B for RP

Reply

[-]

xor_2@reddit

Why would you say that? This model identified serious issue from simple 'hi' - cannot get more bang for buck!

Reply

[-]

PavelPivovarov@reddit

It is outdated but it's still kicking hard. Probably the most playful model I've ever tried and it's also a joy to come back to it from time to time. I haven't tried LR variant as on the picture but original tiefighter is still amazing, even comparing to Rocinante, Niitama or Stheno. I specifically like it's ability to pick up the conversation without any prompt. Just drop to it something cheesy and it will instantly align with it, or call it "baby" and it will play the partner role right away, never saw anything like that elsewhere. But you're right it's old by now.

Reply

[-]

Fusseldieb@reddit

Are there any worthwhile models for 8GB VRAM poor users?

Reply

[-]

TheDreamWoken@reddit

Why Nemo over qwen2.5

Reply

[-]

nmkd@reddit

That's 13b for ya

Reply

[-]

Hopeful_Ad6629@reddit

Taking "you had me at hello" to a whole new level!

Reply

[-]

DataScientist305@reddit

100% accuracy

Reply

[-]

razziath@reddit

lmao

Reply

[-]

AD7GD@reddit

Reminds me of the time I accidentally set Alexa off on a long philosophical rant. I had to check the phone app to find out it thought I asked it, "What is?"

Reply

[-]

Odd_Sentence_2618@reddit

Mine is barely able to play a simple playlist. Must have been lobotomized too in the process.

Reply

[-]

bullet_proof-monk@reddit

Sure wasnt it loaded in between chat history

Reply

[-]

dagerdev@reddit (OP)

Not at all. I downloaded the model,new chat, said hi. And that's it.

Reply

[-]

OriginalPlayerHater@reddit

lm studio has options for system prompts, cute trick though

Reply

[-]

dagerdev@reddit (OP)

The system prompt was empty. This was pure AI hallucinations. That's why I found this funny.

Reply

[-]

bullet_proof-monk@reddit

LLMs can sometimes hallucinate,meaning they generate outputs that are not based on the input or the training data but are instead random or nonsensical. The LLM’s response is likely due to a combination of statistical associations in its training data, its design (e.g sensitivity to emotions), and the inherent limitations of how these models learn. It’s a good example of how LLMs can sometimes produce outputs that seem intellignt but are actually based on superficial patterns and lack true understandng

Reply

[-]

RehanRC@reddit

peak comedy

Reply

[-]

djstraylight@reddit

The other problem with this model is that it misses the stop tokens sometimes and just hallucinates more & more. Like your grandpa saying, 'did you say something?' and just blathering on.

Reply

[-]

Krigen89@reddit

What did you say about my grandpa?

Reply

[-]

NickNau@reddit

it's not a problem, it's a feature 😅

Reply

[-]

JohnnyLovesData@reddit

https://preview.redd.it/j2wolbrr28je1.png?width=1080&format=png&auto=webp&s=b9f1aac67b03a48126c03891d2244b5618cb8a14

Reply

[-]

MoffKalast@reddit

Purely psychosomatic? That boy needs therapy.

Reply

[-]

ZachCope@reddit

That boy don’t need a shrink he needs a social worker.

Reply

[-]

kalokagathia_@reddit

It's frontier psychiatry!

Reply

[-]

vinson_massif@reddit

HAHAHA!! this made my day. love calvin and hobbes randomly showing up in my life

Reply

[-]

witheringsyncopation@reddit

Upvote for the unexpected Calvin and Hobbes reference

Reply

[-]

CharlieBarracuda@reddit

"You are an unfit mother. Your children will be placed in the custody of Carl's Jr."

Reply

[-]

Stefroooo@reddit

It’s not what you said it’s how you said it.

Reply

[-]

LSeww@reddit

Sounds accurate, no sane person would say "hi" to an LLM.

Reply

[-]

NickNau@reddit

hell yeah! I still keep old rusty TieFighter-Holodeck-Holomax-Mythomax from our bro /u/DavidAU. the day comes and you feel bit unhinged and you know which of them to launch 😅

Reply

[-]

martinerous@reddit

And TieFighter was quite good with following lengthy instructions with dozens of "Then do this" steps. Until I discovered Mixtral 7x8B which was even better.

Reply

[-]

Mmmrrr_donuts@reddit

Thought criminal!

Reply

[-]

LosEagle@reddit

It's not about what you wrote but about the way you wrote it. It was a hi as if it came from an empty void. A shell of the happiness you once embodied.

Reply

[-]

anshulsingh8326@reddit

Get well soon

Reply

[-]

MandateOfHeavens@reddit

Idk, but that "hi" had a lot of repressed anguish behind it. Like, shit dude, you alright?

Reply

[-]

Skymt1@reddit

Yeah, the lack of punctuation is troubling enough, but not capitalizing the h is definitely a sign of having given up on life.

Reply

[-]

tmvr@reddit

It must have been the tone in which you said it.

Reply

[-]

pip25hu@reddit

You may think you only said hi, but the way you failed to capitalize the initial "h" spoke volumes. :P

Reply

[-]

gangofminotaurs@reddit

OP's happier now. He's with the great llm in the cloud.

Reply

[-]

LoafyLemon@reddit

Was it wrong? :P

Reply

[-]

ryanknapper@reddit

It *knows*, and you know what it knows.

Reply

[-]

Confident-Aerie-6222@reddit

you chose the wrong preset. it should use llama2 chat template preset to work properly

Reply

[-]

cdrfrk@reddit

It's lupus

Reply

[-]

first2wood@reddit

Con artist LLM,

Reply

[-]

elswamp@reddit

You have a system prompt that is directing the llm

Reply

[-]

dagerdev@reddit (OP)

There was no system prompt.

Reply

[-]

DangKilla@reddit

I get these responses, too, when I use "Hi" as a test prompt. I think it's a side effect of being trained on LLM chat data. That's the only thing that makes sense to me.

Reply

[-]

darth_chewbacca@reddit

> There was no system prompt. exactly. no system prompt is a clear display of apathy, apathy towards your future, your relationships, your physical health, your financial situation. The LLM knows what it's doing. PS. Hi.

Reply

[-]

admajic@reddit

Using llama 3.1 8b and is flying along

Reply

[-]

ZynthCode@reddit

My condolences - hope you get the help you need :'3

Reply

[-]

llama-impersonator@reddit

ancient merge model. when used with almost no context those l2 merges were pretty much like asking for random incoherence for output.

Reply

[-]

darth_chewbacca@reddit

Dont do it OP! Things will get better.

Reply

[-]

evi1corp@reddit

Did you get the bill yet?

Reply

[-]

SathwikKuncham@reddit

I think you have given a pre-prompt with this context.

Reply

[-]

Massive-Question-550@reddit

To be fair you gave a rather colorful LLM not a lot to work with.

Reply

[-]

ThenExtension9196@reddit

Dafug is in that system prompt lmao

Reply

[-]

ThenExtension9196@reddit

System prompt bro.

Reply

[-]

daredevilthagr8@reddit

Dr. house when someone says hi

Reply

[-]

TwoWrongsAreSoRight@reddit

Wow..i need this as my therapist. Save me hours is useless yammering

Reply

[-]

terminoid_@reddit

damn, tiefighter? blast from the past right there

Reply

[-]

No_Success3928@reddit

loved those games!

Reply

[-]

fotiro@reddit

hi

Reply

[-]

PriceNo2344@reddit

Hope you can get the help you need OP

Reply

[-]

PwanaZana@reddit

Lol :P

Reply

[-]

AaronFeng47@reddit

Don't you know Meta Llama can read your thoughts? You can't hide from Zack https://preview.redd.it/76xfmm8gx7je1.jpeg?width=575&format=pjpg&auto=webp&s=783920e65d8e43731f4f54e26bd2627b9ebda605

Reply

[-]

AaronFeng47@reddit

Llama can read your thoughts

Reply

[-]

oodelay@reddit

Stop denying

Reply

[-]

HSLB66@reddit

I think your LLM might be crying for help. It’s clearly suffering from severe schizophrenia hallucinations

Reply

Reply to Post

76 Comments