great video, love the well structured and informative nature they have. Would love to see how to use word embeddings from Llama2 or other language model for transfer learning. Thanks and keep up the good work!
I am struggling to have llama 2 working with Dutch language reliably, so you can pose questions in Dutch and have llama 2 give the answer in Dutch. (This will be due to the fact that llama 2 is trained on data that contains very little Dutch language). I have had some succes using special prompts to do that, but sometimes it switches back to English unexpectedly. What technique(s) can I use to solve this? My use case is: I have Dutch texts that I want to be able to pose questions to in Dutch by means of Retrieval Augmented Generation (RAG) (using a llama 2 LLM) and get answers in correct Dutch?
Hey, thanks a lot for your videos. Your video - transformer attention is all you need helped me build an intuition back before transformers were really cool. It's lovely to see your video on llama, as I actively get to finetune llama on day to day basis :) Much love.
The more i watch videos , the more i understand a subject, this is propably because i Can now see the subject in different angles or perspectives, now i have a better intuition of transformer architectures and i Can code it from scratch, thank you.
Thank you for such an insightful video. Would definitely love a deep-dive video on the architecture and code of LLama 2. Could you please also do an implementation of BERT or RoBERTa fine-tuning (the training process optimized via deepspeed) . Thanks again!!
I think you didn’t describe RLHF fully. What you described was more SFT, you seemingly skipped mentioning the reward model explicitly. Maybe implicitly you meant it, but it could help to clarify this part of reinforcement learning
He has a very good series called "chatGPT explained" where he goes into detailed explanation of RHLF: ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-_MPJ3CyDokU.html
Yea that’s true. I mentioned this as “humans determining what is a better answer” when I probably should have said “humans determine the better answer to train the rewards model (s) and this in turn is used with the original fine tuned model to further fine tune it. And this happens via some proximal policy optimization” ~ or maybe something along these lines. Thanks for pointing it out. I’ll clarify this on some follow up videos in the near future too
I have not implemented the code for decoder only so I have 3questions: 1. so it uses the triangular mask? I have heard from 2 sources which it does, but I dont get it, as we only feed inputs and not the outputs(unlike original transformer),how triangular mask on input data makes sense? 2. does why its called `decoder only`? the architecture seems much closer to encoder part of original transformer model, than its decoder part!! specially when the mask also not different than encoder of original. 3. is it autoregressive or still can be autoencoder to output the outputs in one pass?
Can you please help me. I have multiple doubts. As all of these models are LLM and these generated next words based on the previous words, can I find tune them on any type of data, for example I like to make a model which can make poems, shayeri for me so can I train these for this task. Also as llama doesn't have an encoder. Isn't it a disadvantage. Also can you please make a video on encoder and decoder and their specific details. Please 🤓🤓
Hi Ajay, would love to hear your insights on PEFT - the theoretical aspects of course. I have seen a lot of videos on PEFT and some reading too. The theoretical aspects are not well explained.
Fine tuning is done depending on the specific task you want. In llama chat and ChatGPTs case, we want the fine tuning on question answering. So we feed the model a bunch of questions + answer pairs and the model parameters are “fine tuned”. Hope this helps.
Hi Ajay, I have been reading Llama 2 research paper. They talk a lot of safety during pre-training as you might have seen. Do you think they score over GPT in this aspect?
Yea. That 77 page dissertation in llama 2 definitely makes the claim that it is safer. They have sections and infographics dedicated to showing this as well. That said, I would need to check how much of this safety is incorporated in the pre training as well. I didn’t think there would be much in this phase. But I haven’t read the entire dissertation, so I may be wrong.