Your articles and videos have been extremely helpful in understanding how LLMs are built. Building LLM from Scratch and Q and AI are resources that I am presently reading and they provide a hands-on discourse on the conceptual understanding of LLMs. You, Andrej Karpathy and Jay Alammar are shining examples of how learning should be enabled. Thank you!
Haha, thanks! I've learned so much thanks to all the amazing people in open source, and I'm very flattered by your comment to potentially be counted as one of them :)
I am your fan, I have most of your books, thanks for this excellent video ! Another evaluation metric that I found interesting in another channel was to make the LLMs to play chess against each other 10 times.
What wonderful Tech Minds : { Sebastian Raschka, Yann LeCun, Andrej Karpathy, ...} who share their works and beautiful ideations for Mere mortal like me... Sebastian's teachings are so, so fundamental that takes fear off my clogged mind... 🙏 Although I am struggling to build LLMs for specific & niche areas, I am confidant of cracking them with great resources like : Build a Large Language Model (From Scratch)!!!
Hi, nice videos! One question for my understanding. When talking about embedding dimensions such as 1280 in "gpt2-large" do you mean the size of the number vector encoding the context of a single token or the number of input tokens? When comparing gpt2-large and Lama2 the number is the same for the ".. embeddings with 1280 tokens".
Good question, the term is often used very broadly and may refer to the input embeddings or the hidden layer sizes in the MLP layer. Here, I meant the size of the tokens that are embedded.
Thanks for the detailed videos and articles. I want to ask if it's possible to create a customized tokenizer as an extension to existing ones for a custom dataset? Also, how do decoder-only models handle other tasks like summarization, and classification after fine-tuning without forgetting their causal pre-trained causal next token task?
Good question. Yes you, can do that, tiktoken for example allows you to extend the vocabulary with additional tokens. However, you have to keep in mind that you'll always have to update the embedding layer and output layer with these tokens in case you want to use the updated tokenizer with an existing LLM. Regarding your second question, you could do that but that would not be ideal because only the last tokens contains information about all other tokens. If you use other tokens, you'll have more information loss.
00:02 Three common ways of using large language models 02:39 Developing LLM involves building, pre-training, and fine-tuning. 07:11 LLM predicts the next token in the text 09:30 Training LLM involves sliding fixed size inputs over text data to create batches. 14:22 Byte pair encoding and sentence piece variations allow LLMs to handle unknown words 16:42 Training sets are increasing in size 21:09 Developing an LM involves architecture, pre-training, model evaluation, and fine-tuning. 23:14 The Transformer block is repeated multiple times in the architecture. 27:22 Pre-training creates the Foundation model for fine-tuning 29:28 Training LLMs typically done for one to two epochs 33:44 Pre-training is not usually necessary for adapting LLM for a certain task 35:51 Replace the output layer for efficient classification. 39:54 Classification fine-tuning is key for practical business tasks. 42:01 LLM instruction data set and preference tuning 45:58 Evaluating LLMs is crucial, with MML being a popular metric. 48:07 Multiple choice questions are not sufficient to measure an LM's performance 52:34 Comparing LLM models for performance evaluation 54:32 Continued pre-training is effective for instilling new knowledge in LLMs 58:28 Access slides on the website for more details
Vielen Dank fuer das Interesse an meinem Buch. So weit ich es vom Verlag mitbekommen habe wird diese Woche zum Drucker geschickt und dann sollte es hoffentlich in ein paar Wochen auch bein Amazon.com/.de als Kindle Version erhaeltlich werden.
Great question. I do think that special purpose ML solutions still have and will continue to have their place. The same way ML didn't make certain more traditional statistics based models obsolete. Regarding deep learning ... I'd say LLM is a deep learning model itself. But yeah, almost everything in deep learning is nowadays either a diffusion model, transformer-based model (vision transformer and most LLMs), or state space model
@16:37 when you say Llama was trained on 1T token, do you still mean there was 32K unique token ? because on your blog post you have "They also have a surprisingly large 151,642 token vocabulary (for reference, Llama 2 uses a 32k vocabulary, and Llama 3.1 uses a 128k token vocabulary); as a rule of thumb, increasing the vocab size by 2x reduces the number of input tokens by 2x so the LLM can fit more tokens into the same input. Also it especially helps with multilingual data and coding to cover words outside the standard English vocabulary."
Thanks for the comment! So in the talk these are the dataset sizes using the respective tokenizer that was used during model training. The vocabulary sizes that the models used are 32k for Llama 2 and 128k for Llama 3.1. So, regarding "do you still mean there was 32K unique token", the vocabulary was 32k unique tokens (but there could be more unique tokens in the dataset). I hope this helps. Otherwise, please let me know, happy to explain more!
I'm interested in SLM RAG with Knowledge graph traversal/search for RAG dataset collection and vector-JIT semantic match for hybrid search. Any repos you think I would be interested in?
Thanks for your interest in this! It's already available for preorder (both on the publisher's website and Amazon) and if the production stage goes smoothly, it should be out by the end of of August
@@alokranjansrivastava623 Architecture-wise, it's kind of the same thing though, except it doesn't have the causal mask, and the pretraining task is not next-token prediction but predicting masked tokens (plus sentence order prediction).
That's a good question. You can do both. By default all tokens, but more commonly you'd mask the tokens. In my book, I include the token masking as a reader exercise (it's super easy to do). There was also a new research paper a few weeks ago that I discussed in my monthly research write-ups here: magazine.sebastianraschka.com/p/llm-research-insights-instruction
@@SebastianRaschka Thanks for the reply I just have a general question: do we use masking? For example, was masking used during the instruction fine-tuning of LLaMA 3 or mistral any Open source LLMs? Also, does your book include any chapters on the parallelization of training large language models?
@@bashamsk1288 Masking is commonly used, yes. We implement it as the default strategy in LitGPT. In my book we do both. I can't speak about Llama 3 and Mistral regarding masking, because while these are open-weight models they are not open source. So there's no training code we can look at. My book explains DDP training in the PyTorch appendix, but it's not used in the main chapters because as a requirement all chapters should also work on a laptop to make them accessible to most readers.
@@SebastianRaschka Hey, sorry for the bad language. I meant should the chat formats like alpaca etc be applied or we give the text as it is to LLM for classification.
@@ArbaazBeg Oh I see now. And yes, you can. I wanted to create an example and performance comparison for that to the GitHub repo (github.com/rasbt/LLMs-from-scratch) some time. For that I wanted to first instruction-finetune the model on a few more spam classification instructions and examples though.
That's great Dr. Hope you will be doing good. I wish if i could download your deep learning book which is published recently. If there is any open source link to download it please mention in comments. Thanks and regards, Researcher at Texas
I've heard about instruct learning, and it sounds similar to how you define preference learning. I have also heard about transfer learning. How would you compare/define those?
Transfer learning is basically involved in everything you do when you start out with a pretrained model. We don't really name or call it out explicitly anymore because it's so common. In instruction finetuning, the loss function is different from preference tuning mainly. Instruction finetuning trains the model to answer queries, and preference finetuning is basically more about the nuance of how these get answered. All preference tuning methods that are used today (DPO, RLHF+PPO, KTO), etc. expect you to have done instruction finetuning on your model before you preference finetune.
@@joisco4394 Btw I recently coded the alignment (using direct preference optimization) here, which might help clarifying this step: github.com/rasbt/LLMs-from-scratch/blob/main/ch07/04_preference-tuning-with-dpo/dpo-from-scratch.ipynb
I have been in this field for a while (decade +) but not in touch with LLMs and it is useless for uninformed and even more useless for the informed. I don't know I am in which category but I didn't learn anything. I read about transformers when the paper came then I assumed RLHF is just reinforce algorithm which is probably correct to assume. Anyways highly repellent video.
ha, I should mention that all chapters in my book run on laptops, too. It was a personal goal for me that everything should work even without a GPU. The instruction finetuning takes about ~30 min on a CPU to get reasonable results (granted, the same code takes 1.24 min on an A100)
Your articles and videos have been extremely helpful in understanding how LLMs are built. Building LLM from Scratch and Q and AI are resources that I am presently reading and they provide a hands-on discourse on the conceptual understanding of LLMs. You, Andrej Karpathy and Jay Alammar are shining examples of how learning should be enabled. Thank you!
Thanks for the kind comment!
You are the best! Thanks a lot for sharing your knowledge to the world.
The hero of open source
Haha, thanks! I've learned so much thanks to all the amazing people in open source, and I'm very flattered by your comment to potentially be counted as one of them :)
Thank you Sebastian for your awesome contributions. You're a big inspiration.
You are a true educator. Honored to be a contributor to one of your libraries.
I know you don't do many tutorials but personally I love theme especially from you!
Thanks, that's very motivating to hear!
One of the best 60 minutes of my time. Really thankful for this..
Thanks for the kind words!
Thank you, Sir. Your lessons are beneficial for the community. Appreciate your hard work..!! 😊
I am your fan, I have most of your books, thanks for this excellent video ! Another evaluation metric that I found interesting in another channel was to make the LLMs to play chess against each other 10 times.
Hah nice, that's a fun one. How do you evaluate who's the winner, do you use a third LLM for that?
u r a LEGEND,luv ur work,thnx a ton for sharing!
What wonderful Tech Minds : { Sebastian Raschka, Yann LeCun, Andrej Karpathy, ...} who share their works and beautiful ideations for Mere mortal like me... Sebastian's teachings are so, so fundamental that takes fear off my clogged mind... 🙏
Although I am struggling to build LLMs for specific & niche areas, I am confidant of cracking them with great resources like : Build a Large Language Model (From Scratch)!!!
Hi, nice videos! One question for my understanding. When talking about embedding dimensions such as 1280 in "gpt2-large" do you mean the size of the number vector encoding the context of a single token or the number of input tokens? When comparing gpt2-large and Lama2 the number is the same for the ".. embeddings with 1280 tokens".
Good question, the term is often used very broadly and may refer to the input embeddings or the hidden layer sizes in the MLP layer. Here, I meant the size of the tokens that are embedded.
Very nice video, I liked it so much that I preordered your new book directly after watching it (to be fair I have read your blog for some time now).
Thanks! I hope you are going to like the book, too!
Thanks for the detailed videos and articles. I want to ask if it's possible to create a customized tokenizer as an extension to existing ones for a custom dataset? Also, how do decoder-only models handle other tasks like summarization, and classification after fine-tuning without forgetting their causal pre-trained causal next token task?
Good question. Yes you, can do that, tiktoken for example allows you to extend the vocabulary with additional tokens. However, you have to keep in mind that you'll always have to update the embedding layer and output layer with these tokens in case you want to use the updated tokenizer with an existing LLM. Regarding your second question, you could do that but that would not be ideal because only the last tokens contains information about all other tokens. If you use other tokens, you'll have more information loss.
00:02 Three common ways of using large language models
02:39 Developing LLM involves building, pre-training, and fine-tuning.
07:11 LLM predicts the next token in the text
09:30 Training LLM involves sliding fixed size inputs over text data to create batches.
14:22 Byte pair encoding and sentence piece variations allow LLMs to handle unknown words
16:42 Training sets are increasing in size
21:09 Developing an LM involves architecture, pre-training, model evaluation, and fine-tuning.
23:14 The Transformer block is repeated multiple times in the architecture.
27:22 Pre-training creates the Foundation model for fine-tuning
29:28 Training LLMs typically done for one to two epochs
33:44 Pre-training is not usually necessary for adapting LLM for a certain task
35:51 Replace the output layer for efficient classification.
39:54 Classification fine-tuning is key for practical business tasks.
42:01 LLM instruction data set and preference tuning
45:58 Evaluating LLMs is crucial, with MML being a popular metric.
48:07 Multiple choice questions are not sufficient to measure an LM's performance
52:34 Comparing LLM models for performance evaluation
54:32 Continued pre-training is effective for instilling new knowledge in LLMs
58:28 Access slides on the website for more details
Ich nehme stark an, dass Du Deutsch sprichst :). Wo kann man Dein Buch im Kindle (mobi oder f2b) Format finden? Danke & LG.
Vielen Dank fuer das Interesse an meinem Buch. So weit ich es vom Verlag mitbekommen habe wird diese Woche zum Drucker geschickt und dann sollte es hoffentlich in ein paar Wochen auch bein Amazon.com/.de als Kindle Version erhaeltlich werden.
@@SebastianRaschka super, vielen Dank.
Oh, my lord, my favourite machine learning author is a Liverpool fan.😎
Haha, nice that people make it that far into the video 😊
@@SebastianRaschka You'll never walk alone 🙂
Great Video. Now that LLM is so powerful , will regular machine learning & deep learning slowly vanish?
Great question. I do think that special purpose ML solutions still have and will continue to have their place. The same way ML didn't make certain more traditional statistics based models obsolete. Regarding deep learning ... I'd say LLM is a deep learning model itself. But yeah, almost everything in deep learning is nowadays either a diffusion model, transformer-based model (vision transformer and most LLMs), or state space model
the greatest Liverpool fan ever! ⚽
Haha nice, at least one person watched it until that part :D
@16:37 when you say Llama was trained on 1T token, do you still mean there was 32K unique token ? because on your blog post you have "They also have a surprisingly large 151,642 token vocabulary (for reference, Llama 2 uses a 32k vocabulary, and Llama 3.1 uses a 128k token vocabulary); as a rule of thumb, increasing the vocab size by 2x reduces the number of input tokens by 2x so the LLM can fit more tokens into the same input. Also it especially helps with multilingual data and coding to cover words outside the standard English vocabulary."
Thanks for the comment! So in the talk these are the dataset sizes using the respective tokenizer that was used during model training. The vocabulary sizes that the models used are 32k for Llama 2 and 128k for Llama 3.1. So, regarding "do you still mean there was 32K unique token", the vocabulary was 32k unique tokens (but there could be more unique tokens in the dataset). I hope this helps. Otherwise, please let me know, happy to explain more!
Thanks for the great knowledge You are sharing
I'm interested in SLM RAG with Knowledge graph traversal/search for RAG dataset collection and vector-JIT semantic match for hybrid search. Any repos you think I would be interested in?
bookmarked, clear and concise.
Unfortunately I don't have a good recommendation here. I have only implemented standard RAGs without knowledge graph traversal.
When is your whole book coming out ? Eagerly waiting 😅
Thanks for your interest in this! It's already available for preorder (both on the publisher's website and Amazon) and if the production stage goes smoothly, it should be out by the end of of August
Nice Video.
Does LLM mean only auto-regressive models (Not Bert)?
Yes, here LLM is basically synonymous with decoder-style autoregressive model like Llama, GPT, Gemma, etc.
@@SebastianRaschka Bert has stack of encoder transformers, but it is not LLM. Am I correct here?
@@alokranjansrivastava623 Architecture-wise, it's kind of the same thing though, except it doesn't have the causal mask, and the pretraining task is not next-token prediction but predicting masked tokens (plus sentence order prediction).
@@SebastianRaschkaJust One question. How to define LLM? When we can say that this particular Language model is of LLM category.
in the instruction fine tuning we propagate loss only on output text tokens? or for all tokens from start to EOS?
That's a good question. You can do both. By default all tokens, but more commonly you'd mask the tokens. In my book, I include the token masking as a reader exercise (it's super easy to do). There was also a new research paper a few weeks ago that I discussed in my monthly research write-ups here: magazine.sebastianraschka.com/p/llm-research-insights-instruction
@@SebastianRaschka
Thanks for the reply
I just have a general question: do we use masking? For example, was masking used during the instruction fine-tuning of LLaMA 3 or mistral any Open source LLMs? Also, does your book include any chapters on the parallelization of training large language models?
@@bashamsk1288 Masking is commonly used, yes. We implement it as the default strategy in LitGPT. In my book we do both. I can't speak about Llama 3 and Mistral regarding masking, because while these are open-weight models they are not open source. So there's no training code we can look at. My book explains DDP training in the PyTorch appendix, but it's not used in the main chapters because as a requirement all chapters should also work on a laptop to make them accessible to most readers.
Should we give prompt to LLM when fine tuning for classification with last layer modification or directly pass the input to the LLM like in deberta?
Thanks for the comment, could you explain a bit more what you mean by passing the input directly?
@@SebastianRaschka Hey, sorry for the bad language. I meant should the chat formats like alpaca etc be applied or we give the text as it is to LLM for classification.
@@ArbaazBeg Oh I see now. And yes, you can. I wanted to create an example and performance comparison for that to the GitHub repo (github.com/rasbt/LLMs-from-scratch) some time. For that I wanted to first instruction-finetune the model on a few more spam classification instructions and examples though.
That's great Dr. Hope you will be doing good.
I wish if i could download your deep learning book which is published recently. If there is any open source link to download it please mention in comments.
Thanks and regards,
Researcher at Texas
Danke!
great content! love it ❤
I've heard about instruct learning, and it sounds similar to how you define preference learning. I have also heard about transfer learning. How would you compare/define those?
Transfer learning is basically involved in everything you do when you start out with a pretrained model. We don't really name or call it out explicitly anymore because it's so common. In instruction finetuning, the loss function is different from preference tuning mainly. Instruction finetuning trains the model to answer queries, and preference finetuning is basically more about the nuance of how these get answered. All preference tuning methods that are used today (DPO, RLHF+PPO, KTO), etc. expect you to have done instruction finetuning on your model before you preference finetune.
@@SebastianRaschka Thanks for explaining it. I need to do a lot more research :p
@@joisco4394 Btw I recently coded the alignment (using direct preference optimization) here, which might help clarifying this step: github.com/rasbt/LLMs-from-scratch/blob/main/ch07/04_preference-tuning-with-dpo/dpo-from-scratch.ipynb
@@SebastianRaschka Much appreciated
Would
you make videos about time series and trannsformer?
Big fan ❤
Im sold , im buying your book .. would love to chat with you sometime if possible
Thanks, hope you are liking it! Are you going to SciPy in July by chance, or maybe Neurips end of the year?
@@SebastianRaschka unfortunately not,but I'd like to have a zoom/google meet chat with you if possible
What we are seeking not only fuzzing code
Glad that's useful
Hello sir, please also make with javascript
I have been in this field for a while (decade +) but not in touch with LLMs and it is useless for uninformed and even more useless for the informed. I don't know I am in which category but I didn't learn anything. I read about transformers when the paper came then I assumed RLHF is just reinforce algorithm which is probably correct to assume. Anyways highly repellent video.
🤌
I take that as a compliment!? 😅😊
@@SebastianRaschka Yes yes! It was supposed to be a compliment only. You are doing great work with our teaching materials :).
Easier said than done unless u got a GPU super computer lying around lol
ha, I should mention that all chapters in my book run on laptops, too. It was a personal goal for me that everything should work even without a GPU. The instruction finetuning takes about ~30 min on a CPU to get reasonable results (granted, the same code takes 1.24 min on an A100)