How to Train a Multi Modal Enterprise Grade LLM from Scratch @ Dolby Laboratorie

Join us for the next AI User Group Meetup where we will discuss the intricacies of what it takes to train an enterprise grade Multi Modal Large Language from scratch.

The meetup is being hosted at Dolby Laboratories where we will also have a demo of Dolby Vision and Dolby Atmos technologies in their state-of-the-art Dolby Cinema.

• ​Dataset Preparation: We'll delve into the intricate process of preparing datasets for enterprise-grade multi-modal language models. This involves the meticulous collection and curation of diverse data sources, crucial for model robustness. We'll explore advanced techniques for cleaning and preprocessing various data modalities, including text and images. Additionally, we'll discuss the challenges and best practices in annotation and labeling, which are fundamental for effective supervised learning tasks.

• Model Training: Our discussion will cover the complexities of training large-scale multi-modal models. We'll examine the nuances of architecture design and the critical role of hyperparameter tuning in optimizing model performance. The conversation will extend to the challenges of distributed training on high-performance computing clusters, a necessity for models of this scale. We'll also explore iterative fine-tuning strategies and rigorous evaluation methods to ensure model quality and reliability.

• ​Inference Optimization: Lastly, we'll tackle the crucial aspect of inference in production environments. This includes cutting-edge techniques for model optimization, such as quantization and pruning, which are essential for efficient deployment. We'll discuss the intricacies of setting up scalable inference infrastructure to handle enterprise-level demands. The conversation will also cover advanced strategies for reducing latency and maximizing throughput, critical factors in real-world applications of these sophisticated models.