"okay, but I want GPT to perform 10x for my specific use case" - Here is how

excellent tutporial! thank you for sharing.

hey man, thanks so much for this. Because of you, I managed to build my first model. Not looking exactly how I want it to, but let me know if you would have some time to troubleshoot.

how do i implement it into my website (to offer it to customers)?!

v good tutorial, that goes through a lot of stuff, from overview, datasets, even dataset creation, downloading, finetuning and even uploading a model; wow

Thanks for sharing such an amazing content! It would be great with you could also make a video about creating embeddings

hey i am getting
Inference tensors cannot be saved for backward. To work around you can make a clone to get a normal
tensor and use it in autograd.
error while running the finetuning part of the code please suggest what to do

Awesome! I'm totally new to fine-tuning. I'm looking forward to fine-tuning a model with thousands of private 4K images!

Now I am trained to subscribe to you and Like all such valuable vids.

Nice, but I think there's a fundamental problem everyone seems to be ignoring, like so many LLM tutorials, either you showed clearly how it isn't possible in practice... or you showed how to build a broken prompt generator. There is an astronomical difference between an automation that is 10x better and 'almost works' vs 'actually works', in fact its binary, one is useful, the other is worse than useless.

Great video, man.

I also want personal ai cause I tried prebuilt ais and they eat the entire dual core of my cpu and heat up my entire house with my 11 year old laptop.

I am not understanding,what he did.
I need to train a model with my custom data,where model need to answer my data related question only. Is it possible by his method?

Hello, im trying to train an ai on a data set listing ratings of professors and what they teach to have the ai answer questions effectively on best student rated teachers for each classes. In this tutorial i see you are using prompts to train the ai, is their a way to just train the ai with the table or do i have to formulate the tables into prompts. Thank you

Why don't you use SFTTrainer? When to use it and when not? Amazing video by the way!!!

This didn't age well 😂

Fastest sub ever...!

if u want much higher level AI guys I'll introducing AI CHAT DEEPAI the AI CHAT DEEPAI is so powerful and much better than OPENAI since start 2015-2018 and u can show ur picture and video and anything u want

Thought it was another clickbait meme video, and here am I, writing a comment about how cool it is, straight to the point, stuffed with info, super interesting, informative and useful in case someone needed to make his language model and was searching for info on it

Imo, way to go is to hard-code it's language processor. Then teach it how YOU do research and find information online. Hook it up with chameleon and train it to take what you say, do the research and extract relevant data from it's web searches. Next, it needs to know what you're using it for and how you want the data. So for programming assistant, it needs to know a programming language and it's structures and use-cases, so it can take the learned data, make it functional (i.e. replace hard values with assigned variables) and retrofit it to your need. Eg. if you want it to make you a calculator app in python, it needs to know english to understand the question, researching, so it can find reputable data to use, python in order to compile the code into something that can run and how you want it to be done (eg. CLI, or pySimpleGUI interface). With those, it'll be able to make you just about anything, provided you're detailed enough on how it needs to be done. Eg. if you're making a large program, you need to include compartmentalization and unit testing, so it can assemble the blocks, before building the final product, so when you want it to make edits, it can do so on the unit test models and swap them in. Alternatively, when you're dealing with financial data, it needs to know best-practices, eg. writing debit before credit, referencing the original and if it's real-time, adding a timer to count how old the data is. Functional ML is very easy, if you have clearly defined goals and requirements.

Great video mate! Keep up the work (Y)

Wow, what a sick video. Thanks so much!

I know you mentioned using embeddings is good for gaining knowledge, but could you use it for creative writing? Could I write the last chapter without the model forgetting previous events to avoid plot holes and conflicting events?

Zwdbwf⁰😅😅😅😅😅😅

Sweet.

I'm commenting before I watch the video because that thumbnail is so god dammed hilarious.
Edit: Now that I've watched this, I want to like it again.

This man is the source.

We just watched the beginning of a literally billion dollar industry appear in less than 10mins

Every day, when you wake up, think that today I am lucky to be alive, I have a precious human life, I will not waste this life

`.to` is not supported for `4-bit` or `8-bit` models. Please use the model as it is, since the model
has already been set to the correct devices and casted to the correct `dtype`.

Can you tell how to solve this error?

Hello, I have question regarding the hardware for LLM transfer learning. If I want to use the pre trained LLM model, and do transfer learning on related domain data. How much system RAM I required? and, how much vRAM required for the gpu? (I want to do the transfer learning only, not want to trained LLM from scratch). Thank you for your valuable time.

dude, blur your background..lol

If Jason grew hair like Jimmy O Yang.....

The Jin Yang trick

how do i add a metric score measuring mechanism like bleu score in the code??

Really useful video, is there a video from you going into detail the difference between embedding and fine tuning in general. The concept is a bit vague to me

Let's be honest. Who here clicked because you thought he was JIMMY YANG?

can gpt read csv documents aswell?

lol, it spend 1 min to generate 1 prompt on so powerful GPU(similar performance as gtx4090).. It looks almost useful in most cases😥

Hi, this was really useful and i managed to finetune a model to my needs. How do I then start using it on it's own?

Duuude it's so funny that you used Jian Yang on your thumbnail cause you guys sound almost exactly alike! Are you imitating him or something?

your strategies of earning money on binary options are interesting. now I'm curious about it

most underrated channel in AI. Thank you for the effort you are putting into this.

awesome video man! thanks a lot :)

Thank you so much for the video! Is it possible to use the collar notebook for another model?

Thank you so much

I saved the finetuned model and pushed it in my huggingface account as your instruction. But when I loaded it, I got the error my model "does not appear to have a file named config.json". Please help me to load it.

Getting this error when running the "Finetune the model" section of the google collab:

Inference tensors cannot be saved for backward. To work around you can make a clone to get a normal
tensor and use it in autograd

I already loaded the data in and every section before this ran without errors. Any idea what the problem might be? I am using T4 GPU