Inside the LLM: Visualizing the Embeddings Layer of Mistral-7B and Gemma-2B
Вставка
- Опубліковано 9 бер 2024
- We look deep into the AI and look at how the embeddings layer of a Large Language Model such as Mistral-7B and Gemma-2B actually works.
You will learn how tokens and embeddings work and even extract out and load the embeddings layer from Gemma and Mistral into your own simple model, which we will use to visualize the model
You will see how an AI clusters terms together and how it can cluster similar words, build connections which cover not just similar words but also grouping of concepts such as colors, hotel chains, programming terms.
If you really want to understand how an LLM's works or even build your own LLM then starting with the first layer of a Generative AI model is the best place to start.
Github
-----------
github.com/chrishayuk/embeddings - Наука та технологія
this is the github repo: github.com/chrishayuk/embeddings
Fantastic video !
I am wondering: I think it would also be very interesting to also be able have a visualization of not only the static embeddings you already did, but also a visualization of the so-called contextualized embeddings in a later layer of the model! These are the embeddings that are exposed to the attention mechanism. That why they are also called dynamic embeddings.
It adds another layer of abstraction, but are better embeddings because they are able to distinguish between homonyms: words that are the same but have completely other meanings if used in another context. A good example is the word “bank”, that has several different meanings when used in another context (for example financial institution or river bank and several other meanings! ). As a consequence the word “bank” will be represented by several different vectors in embedding space, depending on the context it is used in!
This technique is called Word Sense Disambiguation (WSD).
Would it be possible to visualize that too? I am curious….
yep, you got what i'm doing... i'm literally walking the stack
so those videos will be coming
@@chrishayukFantastic ! Those embeddings are crucially important for the workings of Large Language Models !
Great insight, thanks for posting this. It would be interesting to show how a fine-tuned model differs in similarities and "vocabulary". I'm also curious on the effects of quantisation, i.e. Q4, Q6, Q8, fp16 etc. on the internal "workings" of the LLM. Thanks again.
It’s almost like you’re reading my roadmap
This came to the absolute right time! Thank you very much! I was just trying to understand this. Now I know how it works ❤
Glad it was helpful!
I was playing a bit with finetuning to force an output schema for some 7B Models, but lately I discovered schema grammar, which is a way to dynamically play with the EOS tokens, by limiting them to a specific set of tokens, to generate the output you want, This is very stable and way efficient for many cases that we may think it requires finetuning, For me it felt like a new dimension to get the model intentions inline, I loved the unique and efficient way you create your videos, So I wanted to ask you if possible to create a video for us about this, I feel it's very important
that's a good shout
Thx@@chrishayuk
Its an great video to understand the internals via the visualization. Thanks Chris.
Thanks the visualization really helped me.
so glad, seeing it at a lower level really demystifies what's going on
Thank you! Great video!
thank you, glad it was useful
Good! Where can I find your programs?
in my github repo github.com/chrishayuk
This helps thanks!
Glad it helped! :)
Thank you so much! Thank you youtube algorithm for showing such a great video!
Glad you enjoyed it!
based