LoRA Fine-tuning Tiny LLMs as Expert Agents
Tiny LLMs have never been ideal for agentic workflows. They lack the ability to reliably generate function calls; however, this isn't due to any real limitation on LLM size. Instead, it's due to the LLM providers' lack of focus on data that provides quality examples of function calling. Because of that, we can fine-tune expert agents from tiny LLMs such as the 1B parameter Llama 3.2 and get incredible results. In this video, we do just that - we take llama-3.2-1b-instruct, Salesforce's xLAM dataset, and Low-Rank Adaptation (LoRA) fine-tuning via NVIDIA's NeMo Microservices, to create our own tiny LLM agent. Thanks to NVIDIA for sponsoring the video! 📌 NeMo Microservices: https://nvda.ws/4mqG2wH 🚢 Deploying NeMo Code: https://github.com/aurelio-labs/cookbook/blob/main/gen-ai/training/lora/nvidia-nemo/deploying-nemo.ipynb 👾 LoRA Fine-tuning Code: https://github.com/aurelio-labs/cookbook/blob/main/gen-ai/training/lora/nvidia-nemo/nemo-lora-function-calling.ipynb 👾 Discord: https://discord.gg/c5QtDB9RAP X: https://x.com/jamescalam LinkedIn: https://www.linkedin.com/in/jamescalam/ #ai #aiagents #artificialintellegence #programming #coding 00:00 LoRA Fine-tuning Agents 01:34 NeMo Microservices 03:19 NeMo Deployment 07:49 Deploying NeMo Microservices 16:54 xLAM Dataset Preparation 26:49 Train Validation Test Split 28:59 NeMo Data Store and Entity Store 34:14 LoRA Training with NeMo Customizer 42:03 Deploying NIMs 47:10 Chat Completion with NVIDIA NIMs 49:47 NVIDIA NeMo Microservices
