Видео

55555555555555555555555

I'm still lost when it comes to models, there seems to be error after error when training with oobabooga

The amount of energy and resources needed to train and develop LLM is sorely overlooked.

Nice explanation, thank you!!

Thanks for the clear and concise explanation, it was perfect.

Thanks a lot for this explanation. How can you even out the errors via the next weights, when you do not now in advance what activation value the weights will be multiplied with?

top notch content

Wow, this incredibly cool for a web inferface and to keep my training going while i wait for for my python to get better 😂 Would love to know about why you used that specific model and what models it can be ised with? I'd love to try it on an uncesored.

Thank you for this. I am sick of seeing click bait or low quality content when Im trying to self educate about llms and get some good knowledge .

Do you have any videos on how to use already existing Loras? I'm running 7-10b Exl2 models, but I have no idea how to use Loras.

wow such a in debt training. I am an IT/Cyber security senior and played with llm's couple of years back. training a statistics model. Your video's helped me a lot to get up to speed again. Thank you for the detailed training.

tks

thanks

You’re literally a genius! I appreciate you taking the time to share the knowledge with us! Exactly what I was looking for… how to create a dataset and in such a well put together video. Thank you

just works with Nvidia ... the title of video is fake: "NO GPU REQUIRED" but the exemplo he used de GPU.

Great explanation. Some questions: When we are quantising and computing the quantisation loss, do we not need to supply some data for it to compute the loss against? If not, how exactly is this loss computed? (surely we need some inputs and expected outputs to compute this loss, is this why all of the weight errors were 0 when you quantised? ) If we do, could this be interpreted as a form of post training, quantisation 'fine-tuning'? By this I mean that we can use domain data in the quantisation process to help preserve the emergent features in the model that are most useful for our specific domain dataset? Thanks!

Finally some freaking great tutorial! Practical, straight to the point and it works!!

I would pay a lot of money for this information, thank you.

Thanks for the walkthrough! So I'm running the utility on an i7 16 GB RAM (no GPU) laptop, but for me it takes close to 10 mins for each response to fully generate. How did you manage to speed yours up? I'm also using Vicuna 7b

hmm... : I would like to be able to : Update the llm , ie by extrracting the documents in a folder , extracting the text and fine tuning it in ? ie : i suppose the best way would be to inject it as a text dump ~ HOW?(Please) ie take the whole text and tne a single epoch only !: As well as saving my chat history as a input/Response dump : single epoch only . Question : each time we fine tune ? it takes the last layer and makes a copy then trains the copy and replaces the last layer ? as the model weights are FROZEN? does this mean that they dont get updated ....? if so then the lora is applied to this last layer esentially replacing the layer ? If we keep replacing the last layer do we essentially wipe over the previous training ?? i have seen that you can target Specific layers ? ... How to determine which layers to target? then create the config to match these layers? Question : How dowe create a strategy for regular tuning without destroying the last training ? should we be Targetting different layers each fine tuning ? Also Why canwe not tune it Live!! ie while we are talking to it ? or discuss with the model and adust the model whilst talking ? is adjusting the weights done by the AUTOGRAD? NN in pytorch with the optimization ? ie adam optimizer ? as with each turn we can produce the loss from the input by supplying the expected outputs to compare with simuarity so if the output is over a specfic threshhold it would finetune acording to the loss (optimize this(once)) ... ie switching between train and evaluation , (freezing a specific percentage of the model )... ? ie essentially woring with a live brain ??? how can we update the llm with conversation , ??? by giving it the function (function calling) to execute a single training optimization based on user feedback ? ie positive and negative votes... and the current response chain ... ie if the rag was used then the content should be tuned in ?? SOrry for the long post but it all connects to the same thingy?

Thank you so much for simplifying this to such extent. Subscribed

Great video! I am wondering if Oobagooga provides a command line way to perform training. I am using a lab server which the web UI is not accessible.

Just found your channel. Excellent Content - Another sub for you sir!

Amazing, loved it

Hi @aemonAlgiz , I am new to Python (and LLMs) and wanted to try creating a dataset from a book as well. However when running the provided code, I got a warning: "Token indices sequence length is longer than the specified maximum sequence length for this model (181602 > 2048). Running this sequence through the model will result in indexing errors Max retries exceeded. Skipping this chunk." (which happened a lot). The new .JSON file was empty. I tried changing the "model_max_length": from 2048 to 200000 in the tokenizer_config from my model, but that only made the warning disappear (but the result was the same). Would love if anyone has a solution to this :)

Thank you, Aemon! I think your channel is underrated.

Token indices sequence length is longer than the specified maximum sequence length for this model (249345 > 2048). Running this sequence through the model will result in indexing errors I am facing this issue, please help for resolution.

can you explain code to convert pdf to json.. i dont know how you doing that.. it's great and thats what we need.. thanks before

Content is good.. but the comedic interruptions are not helping much, pls

where are you buddy! why did you ditch your youtube channel, this is some of the best content available on the intricate topic of advanced LLM fine tuning and such, please come back to us 😆 where are you , we want new content about Openrouter, Claude, etc' !

great explanations thanks a lot for your efforts making this great content!

MAKE MORE VIDEOS

I'm completely new to ML and LLMs and found this incredibly insightful. Love this type of content.

What is the moat for Nvidia? Can you make a video on that?

what if you have 500 pdf`s you want to use as data

Thank you so much for this video! It was incredibly clear and the concrete example you provided really helped to understand. Your content stands out, and it's greatly appreciated!☺

I came up to an idea (similar to your "crackpot" idea) on my own a week ago, which involves "giving the memories to neural networks" based on my knowledge & some research in biology (even though i'm not a biologist myself) and hacked it into pytorch to test on some pretrained models. I'm still tinkering with this concept of "memories", it requires some tuning of how much the "memories" influence "future state" of the model and which parts of layers to apply the "memory hack" to.. but it works (kind of..). There are some cases when the model goes a bit "crazy", being "overwhelmed" with past "memories", thus i'm still playing with different numbers. Yet, I can't test it on huge models, because I'm just a random guy with some linux & programming skills, and i'm running my tests on a 14y old cpu with 8GB ram 😂

3:00 This is a misrepresentation of softmax. It assigns proportionally more weight to bigger values - that's why it has "max" in its name.

I get "can only concatenate list (not "Tensor") to list" and "Error None"

great job!

Did you merge the qlora with the quantized model or the base model? Is it possible to train on the quantized version and apply it to the base version?

thank you

Dude seriously your content is so clear and easy to follow keep it up!

Greetings. Is everything ok at your end ? We haven’t heard from you for a while and just checking in. Anyway hope you are alright.

I know this video is almost a year old, but I am having some trouble with any of the models I know of. I have tried many different models and there is always an error here. WARNING LoRA training has only currently been validated for LLaMA, OPT, GPT-J, and GPT-NeoX models. (Found model type: LlamaCppModel) WARNING LoRA training has only currently been validated for LLaMA, OPT, GPT-J, and GPT-NeoX models. (Found model type: NoneType) Which models do you use? Does this tutorial still work?

"big O notation" is not the same thing as worst case. You take an input distribution and you can consider average/worst case over that distribution. O gives an "upper bound" to the complexity of that case. If you take quicksort with uniform input distribution, one would typically say that the worst case is O(n^2) and the average case is O(n log n). You could equally well say that the worst case is O(n^3) although this gives you less information.

It's interesting ... still a bit in the dark as to what goal is expected from those inputs and what it does in the final results. Is the goal to style the answers in a certain way (which is how the formatted input seems to focus on) or is the goal to supplement content (a correct Turing biography)? If content is involved, can we train a model to follow a fictional lore, get characters to express behaviors that appear "normal" to its world, or redirect behavior (if someone speaks ill of the king, characters will always respond negatively - or - reacting according to political alignment). And if it's possible, is a large explanatory text like the Turing biography a valid input format to do so ?

using rtx3070,run Lora with ur data only in 5 minutes not hours! but inference is time consuming.Thanks! professor! and I'll be your fans.

do you have a video on how to prepare a dataset for creative writing?