Rubem Didini’s Post

One more open source LLM - DBRx - A New State-of-the-Art Open LLM ! 🌟 https://lnkd.in/gVNuCiJi Trained on 3072 NVIDIA H100s for 90 days - Demand for H100s going High!!! Model parameters: 132b Active parameters: 32b 💡 Despite having 132 billion model parameters, DBRx utilises MOE to efficiently utilise resources, with only 4 experts out of 16 experts used in inference, resulting in active parameters being just 32 billion! 🤯 #DBRx

How I fine-tuned an LLM to write like me! The latest trend in AI is Large Language Models (LLMs) and their ability to generate text, images and even code. I wanted to try one out and see how it would perform. I started by using the excellent and open-source tool Hugging Face’s Transformers For NLP, which allows you to browse, download and use pre-trained models. I downloaded one of the most popular models - LLaMA - and used it to generate some text. It was pretty good, but not great. It lacked the nuance and personality that I try to bring to my posts. So I decided to fine-tune it! Fine-tuning an LLM is a simple process. You need to have a pre-trained model and some Python code. Here’s the code I used: https://lnkd.in/dJ5YHSpK It’s still not as good as my writing, but it’s pretty close! I’m sure that with more fine-tuning it will get even better. What do you think? Have you tried anything similar? -------------------------------------------------- Talk to me about AI, NLP, Machine Learning and Data Science! 👨💻 I’m Mikkel Jensen, a Data Scientist and AI consultant. #AI #NLP #MachineLearning #DataScience -------------------------------------------------------------------------------------------------- This is part two in a series meant to showcase anecdotal evidence, for using an LLM's context window versus fine-tuning it. Link to the first post in comments 👇 The text above was created using the fine-tune version of Llama-2, tuned on my old LinkedIn posts. I changed 2-3 minor things from the prompt output. Find the original output in the code! Things I notice from fine-tuning: Hallucination seems to be a big problem Seems to resemble my style of writing much better than when using the context window Next up: Fine-tuning with more prompt engineering. Feel free to steal the code!

You want to learn how to fine-tune a Large Language Model (LLM) on a consumer-type hardware (e.g. free-tier Google Colab instance)? Check out how to do it, through a step by step tutorial in this recent blogpost from PyTorch: https://lnkd.in/e5jgvein Original post: https://lnkd.in/e9-DTvC4

Got GPU? https://lnkd.in/gRapK2AR Learn how to put your GPU to work at #THETACON Visit - Thetatoken.org or Thetacon.org to learn more

If you'd like to accelerate your model's inference performance, follow along Het Trivedi's solid guide to deploying LLMs into production using Nvidia's open-source TensorRT LLM framework.

Merge Large Language Models with mergekit Model merging is a technique that combines two or more LLMs into a single model. It's a relatively new and experimental method to create new models for cheap (no GPU required). Model merging works surprisingly well and produced many state-of-the-art models on the Open LLM Leaderboard. Blogbeitrag auf HuggingFace https://lnkd.in/eRgEJvmq

Simple explanation of RAG from NVIDIA blogs….

If you'd like to accelerate your model's inference performance, follow along Het Trivedi's solid guide to deploying LLMs into production using Nvidia's open-source TensorRT LLM framework.

If similar to me some of you are trying to fit LLM fine-tuning experiments onto a single consumer GPU have a look at a short and simple article I wrote on this topic #LLM #finetuning #GPU #GPUexperiments #LLMoptimization #singleGPU #machinelearning #deeplearning #neuralnetworks #techarticle #GPUperformance:

In a comprehensive technical guide, Chaim Rand zooms in on model-training approaches, and discusses "one of the more advanced optimization techniques — one that sets apart the true rock stars from the simple amateurs — creating a custom PyTorch operator in C++ and CUDA."