Young Larry Anderson Sandra Thompson Cynthia

great pres, is there a chance i can get the slides.

Thank you! This was a helpful overview! It helped me to better understand .autolog() especially.

Is it like weights and biases?

Harvard is not known for producing significant AI research. I get Stanford and Berkeley but Harvard?

That was a remarkable presentation. I am not surprised that more Ph.D's are headed into the industry than ever. Some of us feel vindicated by that trend🙂👍🏾. That said, this moment is revolutionary and it is palpable; it is about impossible to ignore the AI storm or merely sleep through it. It is both exciting and a little bit scary. I am personally excited about the possibilities and the repertoire of tools available to humanity to make gigantic innovation leaps in just about every domain. Thanks FourthBrainAI, thanks panel for the insights and perspectives.

Such an amazing and informative talk, thank you!!!🙏

😍

Wow, you are awesome!

Thank you!

Thank you for uploading this :)

hello i need quick help her the env that i set up in WSL dont show in VS Code can u help in this ... ?

Can I get an acess to slides?

Loved it

Please share the jupiter note book or the code...

please turn on the subtitles

Thanks ❤❤❤❤

Can this be done using LORA?

Link to the notebook: colab.research.google.com/drive/1JBtIiMA-LLCmqxGwzK6aokPKR6wRWdV0?usp=sharing

This is the best and most extensive explanation of RAG that I have seen on KZbin. Thank you!

Enhance your marketing efficiency and financial success using Phlanx's Caption Generator, an AI solution that simplifies the content creation process, allowing businesses to focus on core activities while reaping the benefits of increased online visibility.

thank you. Just out of curiosity, can you share code/notebook as well?

A few of the links in the "summary" above result in "page not found". Also is there a link to the slides? Thanks for the great presentation btw. I always enjoy learning the cutting edge of time series forecasting.

This is exciting! I cannot wait to try it out on sensor data. Thanks folks.

you say Time GPT is open source and it is not downloadable - that is not open source - hahaha. What a waste of time - why cant you be clear

multivariate timeseries is required for stock prediction. any plans to extend timegpt to multivariate?

Amazing video and keep it up good work 🎉

awesome! thanks a lot.

when the next MLOps program is scheduled for?

when does the next MLE program begin?

Can u help me? I have a dataset with images of size 1024x536, can I change the input size for fine-tuning?

Excellent session, Salwa and Luis. Thanks FourthBrain 👏👏👏

What resources do you use to finetune the model?

Now I receive "Out of memory on GPU" every time, even tho I use Colab Pro. Maybe someone have any insights on this issue?

I've upgraded colab, and now I've 15 GPU RAM. and receiving this error: torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 50.00 MiB (GPU 0; 14.75 GiB total capacity; 13.29 GiB already allocated; 6.81 MiB free; 13.44 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF However I've make the --max_train_steps=10 Any idea what's the minimum required RAM, and how can I optmize that?

how bout finetuning for video?

A tutorial on using Dreambooth as well would be super cool

Thanks for sharing - it was helpful

🎯 Key Takeaways for quick navigation: 00:00 🎉 Introduction to fine-tuning LLMs with Ludwig framework. 00:27 📚 Learn to fine-tune large language models (LLMs) using Ludwig framework. 00:55 🧠 Ludwig uses a declarative programming approach to generate fine-tuned LLMs with minimal code. 02:16 🛠️ Speaker Piero Molina introduces Ludwig, an open-source deep learning framework with declarative capabilities. 03:36 🚀 Ludwig simplifies training and fine-tuning LLMs, reducing coding effort and time. 07:08 🔧 Ludwig configuration specifies inputs, outputs, and task types to create LLMs with desired architectures. 09:43 🚀 Fine-tuning LLMs with Ludwig requires minimal code - about 10 lines - to achieve effective results. 13:44 📊 Ludwig supports large LLMs like GPT-3, and handles distributed training with Ray for models exceeding GPU memory. 20:28 📊 Evaluation and visualization tools in Ludwig help analyze LLM performance. 23:06 🔄 Switching LLM models or transfer learning with Ludwig is straightforward by adjusting configuration parameters. 25:30 🚀 Developing machine learning models using traditional approaches can be time-consuming and resource-intensive, often requiring months of development and deployment. 26:24 🛠️ Aggregative interfaces and declarative approaches, like those used in tools such as DBT and Terraform, can simplify complex tasks in data engineering and infrastructure management. 27:18 ⚙️ Combining declarative and automation approaches in machine learning reduces time to value and opens opportunities for engineers with varying levels of ML expertise. 28:41 🧠 Ludwig's architecture revolves around ECD (Encoder-Combiner-Decoder), allowing flexibility in encoding different data types for various machine learning tasks. 29:48 🧩 Ludwig's configuration-driven approach lets you build different machine learning models by specifying encoders, preprocessors, architectures, and more. 30:44 📊 The flexibility in the configuration file allows creating multi-modal, multi-task models, enabling tasks like image captioning, regression, audio processing, and more. 33:57 📝 Ludwig's configuration system can be easily extended with custom encoders, expanding the platform's capabilities for various applications. 37:56 ⚙️ Scaling Ludwig is facilitated through the Ray backend, enabling data parallelism and model parallelism for larger data sets and models. 43:45 🌟 Pretty Base enhances Ludwig with additional components, making it an enterprise platform for easy model building, iteration, deployment, and collaboration. 46:04 🌐 Ludwig supports various LLMs, including those available on Hugging Face, and upcoming versions will include built-in support for models like alpaca and llama. 48:47 🏁 When fine-tuning LLMs on resource-limited platforms like Google Colab, consider using smaller versions of models and optimizing for performance per available resources. 50:54 🧠 Fine-tuning LLMs can achieve comparable performance to full fine-tuning or even better in some cases, depending on data similarity. 51:34 🔄 Freezing specific parts of a pre-trained model can be effective when the new data is similar to the original model's training data. 53:51 🏆 Hyperband with Bayesian optimization is recommended for hyperparameter optimization due to its efficient resource usage. 55:29 🌟 Keeping up with AI advancements is challenging, and following influential lab discussions or seminar series can help. 57:21 🚀 Beginners aiming to add ML capabilities to applications can start with high-level tools like Ludwig. 58:30 🔍 For researchers, exploring historical AI papers can provide valuable insights into the field's progression. 01:00:08 🧩 Building with high-level abstractions like Lego blocks is crucial in the evolving landscape of AI. 01:00:47 🤝 Joining AI communities like Stanford MLSys can help stay updated and connected within the field. Made with HARPA AI

Can you provide the fineruning code? Thank you.

Thanks for the video, it has been really useful! Do you know how to keep trace of the training logs?

amazing

Forgot one cut there 22:11 ahah Nice work tho thank you

Hi, I tried navigating to the github url visible in 1:49 timeframe of this youtube video and I dont find any such repo. Can you please share the correct repo url here?

How to join your meetings?

In the second approach did you not use OpenAI API ?

bro you're amazing!

Nice work.

Good video Chris

what is the loss function here, what is the model learning during training?