This is probably most complete and clean LLM lecture in the entire internet. Thank you for your efforts and knowledge I'm currently using the GPT-4 API, which provides extremely accurate results for my project, but it can be expensive if used excessively. As a cost-saving measure, I'm considering using one of the downloadable models you recommended. However, I'm wondering if it's possible to train these models with custom data. Please keep in mind that my computer has limited resources, and the most I can use is Colab+ (the enterprise version is not an option for me).
Everyone one has exact same concern with the OpenAI API cost, however none (in my opinion) of the open source LLM models are as performant and accurate as OpenAPI GPT-x models. Others will contradict with my above statement however for certain business use cases GPT-x models are still top so I understand what you say.. Please try GPT4ALL and CrebrasGPT as open source models to see if they help you out www.cerebras.net/blog/cerebras-gpt-a-family-of-open-compute-efficient-large-language-models/ Thanks for the comment, appreciate it sincerely.
@@650AILab Yes, seems like training a custom LLM model requires tons of expertise, resources and money (after checking the link, realized that training from scratch services may go even over 500K USD). I better stick with gpt4 for now 😅 and wait for your videos if anything new happens Thank you for the clean explanation and informative link
@@emrahe468 Please check out my latest video released today which is on the same topic, training or turning LLM for enterprise - ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-5devEOcgrG0.html Thanks.
Thanks for your comment, it's wonderful. Please remember here we are talking about how to use LLM based on the models are available, so the general tech stack and the content we talked here are different.
@Prodramp that's what I meant, I'm trying to host local llm conversational bot on my local machine, but chat gpt was the only thing I could ask questions too. It even broke it down to 40 key steps beginning with what Linux packages I needed
Thanks for the comment, appreciate it sincerely. You need to separate training from fine-tuning. In the training you have full control and the updated model will have all the trained weight which you can analyze as the way you want. The fine-tuned data is converted into embedding vector which can be stored in the vector storage of your choice.