Dataset Generating App with NVIDIA NIM and Llama 3.1

Рет қаралды 17,196

Күн бұрын

Пікірлер: 51

@PatrickHoodDaniel 3 ай бұрын

This is pretty cool. I have a question, is there a way to gather a bunch of data from a specific person (i.e. the writings of my father that passed away a while ago), and a picture of him, and create an avatar and be able to train an LLM to exhibit his persona? Is this possible yet? I watched a presentation from NVidia on the avatar, but the persona from writings from a specific person I'm not sure about.

@PythonSimplified 3 ай бұрын

Great question!! There's an entire section dedicated to digital avatars in the NIM API Catalogue. Some of it is still in a preview stage, but much of it is already live! I haven't had a chance to check it out yet, but we're definitely approaching the point in time when it would be possible! 🙂🙂🙂 You can have a look here: build.nvidia.com/explore/discover/digital-humans Also, I wonder if somebody else in the comments has more experience with that? Please feel free to pitch in, folks!

@antor44 3 ай бұрын

I have not tried Nvidia NIM and don't know if it supports user's fine-tune, but its Avatar option is divided into three tasks: voice, text reasoning, and video image. All these tasks are performed by retraining each chosen AI for a specific specialization or a very particular case. This retraining is commonly known as fine-tuning in the context of text-based AIs for chatting. In other AIs, fine-tuning or something similar is referred to as LoRA, or there may be other terms depending on the technical modifications used. Generally, only a few example data files are needed for fine-tuning, which is more than sufficient for many cases, and it is not particularly complicated. Basically, fine-tuning takes advantage of the data used to train a large AI model-since, for example, language or many expressions are common to all people-and only adds the specific nuances desired in the new model. There are several options for configuring fine-tuning, and before starting the training, the training texts must be formatted. For instance, the format will differ if the text is meant for a chat, writing short texts, or providing a response or rating based on a single word. It is not clear how much text is needed, as it depends on many factors, such as how different it is from what has already been trained and the quality of the new texts for fine-tuning. In the field of AI programming, quality refers to how different the texts are, and having more texts is generally better. Additionally, the file type and format of the training texts must correspond with those used by the programmer for fine-tuning or the chosen online service. For example, in a typical medical application used to diagnose diseases through a chat with the user, fine-tuning might involve numerous questions from doctors and responses from patients in real cases. The programmer must first separate each "doctor question and patient response" pair found in the training text document or file. Then, these text pairs must be formatted according to the file type specified by the fine-tuning online service or the fine-tuning code if executed locally, as each can differ. They can be JSON files, which are text files with tags indicating they contain the training text, or other tags defining their characteristics as required by the fine-tuning code. In the case of fine-tuning with raw texts or long texts formatted without much attention, this would be the easiest scenario, apart from the other easy option of formatting or splitting them automatically using code. However, this does not offer the same quality as human supervision. On the other hand, manual formatting and supervision by humans may involve a lot of work and does not guarantee the desired results.

@PythonSimplified 3 ай бұрын

@@antor44 Thank you so much for sharing your incredible insights!! I agree that manually curated data is always preferred quality-wise and that there's no easy way of knowing if you're wasting your time or if you're actually producing meaningful samples. Prototyping helps, but often the model will not perform well without a sufficient number of samples. I've encountered it recently with fine-tunning VGG-19 and AlexNet on a dataset I manually curated. In the prototyping stage, I got horrific results (the same class was predicted for all samples), but now when I have way more data - it suddenly works as intended!! 🙃 Thanks again for taking your time and sharing your experience! I hope it will help many folks (myself included) on their LLM fine-tunning journey!! 😀😀😀

@zulqurnainhafeez1257 3 ай бұрын

Hello watching from Pakistan. Your teaching style is phenomenal i quite learning python almost 1 time but then i watch your video and conditions are totally changed it gives me good understanding and new motivation to start again. Thank you, keep growing.

@PythonSimplified 3 ай бұрын

Thank you so much, dear!! Super happy to help and motivate you on your exciting learning journey!!! 😀😀😀

@zacharyjohnson7192 3 ай бұрын

Always an awesome job! Love the thumbnail! Smiling lama is back but in the data centre!

@PythonSimplified 3 ай бұрын

Thank you so much, dear! 🙂 I'm trying the new "test and compare" feature for thumbnails, it shows you guys both versions and it tells me which one worked better! Right now outdoors lama is at 45% and the data center lama is at 55% so it looks like we have a winner! 😉

@alexandermuir8160 3 ай бұрын

Mariya is back. What an awesome video. Even I understood it.

@PythonSimplified 3 ай бұрын

hahahaha Thank you so much, Sandy! Super happy you enjoyed the tutorial! 🙂

@FranksWorldTV 3 ай бұрын

Awesome video!

@PythonSimplified 3 ай бұрын

Thank you so much! Glad you liked it! 🙂🙂🙂

@richardmilian9959 3 ай бұрын

I like the way you teach this technology, the microservices are interesting and as a beginner i need to get familiar with Docker and i'll watch the tutorial where you explain about it to go into more detail. From El Salvador thank you so much for this tutorial and the others that you have created.

@nikluz3807 3 ай бұрын

Nice to see you finally got a pet llama

@PythonSimplified 3 ай бұрын

One step closer to officially becoming a farmer! 🤣🤣🤣

@mauricio9581 Ай бұрын

Can anyone use NIMs? I am interested in Riva but it looks like I need an Enterprise License?

@feed730 3 ай бұрын

beautiful video

@Musikfueralle 3 ай бұрын

Interesting and very good presentation - thank you

@smalirizvi8026 3 ай бұрын

First message!!!! 😊😊 Mariya. How are you?

@PythonSimplified 3 ай бұрын

Hi Ali, super happy to see you in the comments!!! 🙂 I'm wrapping up my BSc and then hopefully I'll have more time to catch up (final project submission on Sep 9th!! can't believe it!! 😱) How about you??

@mrkoge636 3 ай бұрын

thanks for the tips

@PythonSimplified 3 ай бұрын

100%!! Enjoy! 🙂

@harishkumaru589 3 ай бұрын

Hi, your tutorials are so easy to understand. Do you have tutorials in Exploratory Data Analysis with Python?

@markmark2961 3 ай бұрын

We need a single released ASAP from 10:38, just say it LOUDER :D

@PythonSimplified 3 ай бұрын

hahahahaha this song has been stuck in my head since the moment of filming!!! It would definitely be hit! 🤣🤣🤣

@markmark2961 3 ай бұрын

@@PythonSimplified Probably you are already aware of it, but Suno can make it a reality :D

@JesseJuup 3 ай бұрын

🔥👍

@JerryN88 2 ай бұрын

Thanks Mariya. Another great video. Just out of curiosity, do you plan on doing anything with the OpenAI Swarm feature that was just announced?

@patekreol974 3 ай бұрын

Super

@wilgarcia1 3 ай бұрын

💙💙💙💙💙💙💙💙

@travisschwartz3397 2 ай бұрын

❤🎉

@Yachid 3 ай бұрын

lol - I waz gonna pick thiz thumbNail

@PythonSimplified 3 ай бұрын

Awesome!! KZbin actually picked it for me! I'm using the new Thumbnail Research feature, where I upload a bunch of options and after some time - it tells me which one works the best 😉

@Babu-so4sj 3 ай бұрын

Love you

@QorQar 3 ай бұрын

Do you run Coda 12 cardsrtx3050-rtx3090-rtx3060 iam ask card supported cuda12

@jonjohnson2844 3 ай бұрын

It got the UK national dish wrong though :D

@PythonSimplified 3 ай бұрын

I always thought it was something like Yorkshire Pudding, but now Google tells me it's Chicken Tikka Masala. WHAT??? 😅😅😅

@Tobs_ 3 ай бұрын

watched on fast forward, came for the llama, stayed for Lara Croft hacking into computers 👍

@PythonSimplified 3 ай бұрын

I knew you'd appreciate my new Lara Croft: GPU Raider - The Cradle of NIM spinoff 🤣🤣🤣

@AntzOutside 3 ай бұрын

@@PythonSimplifiednerd

@AntzOutside 3 ай бұрын

Btw just joking, your videos are great

@diegoarmandourreamendez1873 3 ай бұрын

I'm second

@Greazvstheworld 3 ай бұрын

Is your channel good for people who know very little to absolutely nothing about computer programming?

@limjuroy7078 3 ай бұрын

Am I the first?

@PythonSimplified 3 ай бұрын

The first human, that's for sure! Apparently 2 bots were faster 😅😅😅

@limjuroy7078 3 ай бұрын

@@PythonSimplified I didn't expect there were bots 😂😂😂 Btw, how will I get the Brev coupon?

@prudhvi6e380 3 ай бұрын

5th viewer

@PythonSimplified 3 ай бұрын

Awesome! Did you end up getting the Brev coupons?? 🙂

@CrispyWafflez-z7x 3 ай бұрын

HI im looking for someone to help me with python. Ik this isnt exactly relevant to the video but i would like it if someone was willing to help me (the creator if possible too) im new and i really want to learn.