So many of your videos really resonate with my experience as a traditional data scientist exploring LLMs. Your "at first I thought prompt engineering was bunk" is definitely my journey as well. I find this to be super highly related to your previous video where you said potentially 95% of use cases can be covered by generalized pre-trained models or fine-tuned models. These models are truly so powerful that the secret sauce is in 1.) choosing the right pre-trained base model 2.) asking it the right questions in an efficient way. Thanks so much for all your work in putting together this content, I find it some of the best-explained LLM content on the interwebs
@user-hv6is9gx6r like using a model pre trained for an appropriate purpose, general purpose models work for a lot, but if I were using a tool to write code, a code specialty model would be better
It's a very nice series! By the way, it would be nice if you considered including examples of using Olama side by side with Chatgpt in your series. I rather use Ollama for testing than ChatGPT
Can you do a video on finetuning a multimodal LLM (Video-LlaMA, LLaVA, or CLIP) with a custom multimodal dataset containing images and texts for relation extraction or a specific task? Can you do it using open-source multimodal LLM and multimodal datasets like video-llama or else so anyone can further their experiments with the help of your tutorial. Can you also talk about how we can boost the performance of the fine-tuned modal using prompt tuning in the same video?
If a programmer builds a working vocabulary and does language design, then prompt engineer as opposed does reverse "engineering" of an existing language in order to find a working vocabulary. The "Fake it till you make it" approach is not usually called science or engineering. So calling this profession “prompt writer” would be more appropriate.
Great work my friend! Can there be a situation where after fine-tuning a model, you still have to do prompt engineering to get the desired output? In other words, can you fine tune a model where one-shot inference works all the time?
While you can always do additional prompt engineering after fine-tuning, it may not be necessary based on the use case. With that being said, no system will ever be perfect. So it is hard to imagine a situation in which one-shot inference will work all the time.
Your videos are great man; I hope your channel grows. Quick question: Langchain seems very integrated with OpenAI's API and software packages; have you tried using Langchain with an open-sourced free of charge LLM? Thanks! I am trying to build an LLM based app for a portfolio for PhD application in AI.
Thanks for the kind words, I'm glad you like the videos. While I've only used LangChain with OpenAI's API, it is has integrations with many other LLM providers. Here's more on how to use it with HF: python.langchain.com/docs/integrations/providers/huggingface
Please explain WHY the correct answer is required in the prompt. I would expect the model to know what the correct answer is. PS: I have enjoyed your other vids and intend on sharing them to my dev friends. Cheers !
Good question. The model does know the correct answer to this particular question. However, there may be questions where the model does not know the answer and providing it in the prompt is necessary.
👉More on LLMs: ruclips.net/p/PLz-ep5RbHosU2hnz5ejezwaYpdMutMVB0
So many of your videos really resonate with my experience as a traditional data scientist exploring LLMs. Your "at first I thought prompt engineering was bunk" is definitely my journey as well. I find this to be super highly related to your previous video where you said potentially 95% of use cases can be covered by generalized pre-trained models or fine-tuned models. These models are truly so powerful that the secret sauce is in 1.) choosing the right pre-trained base model 2.) asking it the right questions in an efficient way. Thanks so much for all your work in putting together this content, I find it some of the best-explained LLM content on the interwebs
Thanks for the kind words. I’m glad you’re enjoying the content. More to come!
@user-hv6is9gx6r like using a model pre trained for an appropriate purpose, general purpose models work for a lot, but if I were using a tool to write code, a code specialty model would be better
It's a very nice series! By the way, it would be nice if you considered including examples of using Olama side by side with Chatgpt in your series. I rather use Ollama for testing than ChatGPT
Thanks for the suggestion :)
Great introduction. Thanks for putting this together.
Glad it was helpful!
Can you do a video on finetuning a multimodal LLM (Video-LlaMA, LLaVA, or CLIP) with a custom multimodal dataset containing images and texts for relation extraction or a specific task? Can you do it using open-source multimodal LLM and multimodal datasets like video-llama or else so anyone can further their experiments with the help of your tutorial. Can you also talk about how we can boost the performance of the fine-tuned modal using prompt tuning in the same video?
Thanks for the suggestion! Multi-modal models are an exciting next step for AI research. I added it to my list.
If a programmer builds a working vocabulary and does language design, then prompt engineer as opposed does reverse "engineering" of an existing language in order to find a working vocabulary. The "Fake it till you make it" approach is not usually called science or engineering. So calling this profession “prompt writer” would be more appropriate.
That's a cool way to think about it. The name isn't great. I can see it being replaced or becoming obsolete.
Great work my friend! Can there be a situation where after fine-tuning a model, you still have to do prompt engineering to get the desired output? In other words, can you fine tune a model where one-shot inference works all the time?
While you can always do additional prompt engineering after fine-tuning, it may not be necessary based on the use case. With that being said, no system will ever be perfect. So it is hard to imagine a situation in which one-shot inference will work all the time.
Your videos are great man; I hope your channel grows. Quick question: Langchain seems very integrated with OpenAI's API and software packages; have you tried using Langchain with an open-sourced free of charge LLM? Thanks! I am trying to build an LLM based app for a portfolio for PhD application in AI.
Thanks for the kind words, I'm glad you like the videos.
While I've only used LangChain with OpenAI's API, it is has integrations with many other LLM providers. Here's more on how to use it with HF: python.langchain.com/docs/integrations/providers/huggingface
Great series! Thanks
Glad you enjoyed it!
Thank you! Awesome content and excellent presentation. Sincerely appreciated 👍
Glad you liked it!
it's really resourceful! keep up the good work
Thanks, glad it helped!
Your acting is ultimate at 1.15 min :)
Thank you 😂😂
Please explain WHY the correct answer is required in the prompt.
I would expect the model to know what the correct answer is.
PS: I have enjoyed your other vids and intend on sharing them to my dev friends. Cheers !
Good question. The model does know the correct answer to this particular question. However, there may be questions where the model does not know the answer and providing it in the prompt is necessary.
So the answer acts like a 'break-glass' test. -- Thanks.
I like the way you present the subject. -- Keep up the good work. -- Cheers @@ShawhinTalebi
Another excellent video!
Thanks :)
Thank you so much
I'm so new at this, but I have to ask...where or which ones are the previous 3?
Here's the series playlist: ruclips.net/p/PLz-ep5RbHosU2hnz5ejezwaYpdMutMVB0
what do the ' \ ' represent in the prompts ? do they break up specific parts of text? Thanks!
Good question. Since the prompt goes over multiple lines, '\' prevents the newline character "
" from appearing in the prompt string.
how can you avoid prompt escape / jailbreak in response?
That's an important (and technical) question. Here is a nice write up on prompt injection: llmtop10.com/llm01/
Fine. I'll roll my eyes less. JK. Great insights on how to improve prompts.
LOL!
I was here for the Sound Effects
I hope it was worth it 😂😂
Tony Stark's part hahahahhahahah
0:58 😂😂😂
😂😂 thanks
@1:03 - they paid you for that didn't they?
That'll be my next career if data science doesn't work out 😂
there is something wrong with you...
LOL what gave it away?