Join us in navigating the practical challenges of working with language models

Language models (LMs) have become increasingly popular for both industrial and academic applications, thanks to their ability to generate high-quality media from textual or visual prompts. Despite their widespread use, working effectively with LMs presents a number of complex challenges that are often underappreciated until deployment. This tutorial aims to explore these challenges in depth, focusing on practical decision-making and implementation challenges when using a LM.

First an overview of popular LM families, such as Gemma, Llama, and DeepSeek, is provided. This is followed by addressing decisions and practices for model selection. The tutorial then covers challenges surrounding training data, which is a key component to specializing a LM for research or application. Creation of training data (extraction from data sources), formatting, and the impact of granularity is discussed. The seven most common prompt styles are examined, with a focus on how to guide behavior through prompt structure. The section on prompts also discusses prompt generation, comparing manual prompt engineering with automated prompt generation. Finally, some of the complexities of working with LM outputs, including evaluation metrics and post-processing techniques.

Throughout examples using in Python using PyTorch are provided in a multi-GPU environment context.