Synthetic Data and Fine Tuning on Databricks

This video is about fine-tuning large language models using Mosaic AI on Databricks! We’ll explore how fine-tuning can significantly enhance the capabilities of smaller models, making them more efficient and accurate for specialized tasks. We'll walk you through the entire process, from generating synthetic data to evaluating model performance, demonstrating how to fine tune LLMs. 💡 Key Moments: [00:10] Why fine-tuning matters: Fine-tuning smaller models boosts efficiency for specialized tasks, reducing costs and improving accuracy. [03:20] Applying fine-tuning to financial regulations: Michael discusses how fine-tuning can help in sectors like finance, specifically for regulatory compliance tasks. [05:45] Data generation demo: Fine-tuning a Llama 3.1 model for European Union capital requirements regulations. The demo shows how synthetic data generation can be done with thousands of questions from PDF files. [12:00] Chunking and cleaning data: Using Spark and Langchain for document chunking and processing, followed by generating tailored questions and answers. [18:30] LLM Judges improvements: Learn how the fine-tuned model’s performance is compared to the vanilla model using LLM judges and evaluation. [25:00] Final thoughts on fine-tuning: Discussion on the ideal data size and strategies for balancing SME input with synthetic data, emphasizing prompt engineering. Link to the repo: https://github.com/mshtelma/techsummit25-model-adaptation-lab/tree/demo #AI #MosaicAI #FineTuning #Databricks #MachineLearning #LLMs #SyntheticData #FinanceAI #ModelOptimization #TechInnovation