Columbia University's AI: Community of Practice Delves into Large Language Models

March 29, 2024

The AI: Community of Practice (AICoP) at Columbia University recently convened for a comprehensive session focused on exploring and applying Large Language Models (LLMs). This gathering sought to clarify the functionalities of LLMs and share knowledge on their diverse applications. Through live demonstrations, participants gained practical insights into the deployment of LLMs, ranging from user-friendly no-code solutions to more advanced, customizable frameworks.

Utilizing ChatGPT Enterprise

The initial segment of the session highlighted the practicalities of developing and customizing conversational agents via ChatGPT Enterprise. Participants were introduced to the platform's no-code capabilities in natural language processing, demonstrating the process of designing bots for specific functions. This involved feeding the bots with domain-specific knowledge and configuring it to deliver precise, reliable responses.

Engaging with Open Source LLMs

A segment was dedicated to utilizing open-source platforms for individuals interested in a more hands-on approach to LLMs. The session explored H2OGPT, a tool for constructing private LLMs. Demonstrations illustrated the process of uploading a JSON file loaded with data from publicly available web pages, showcasing how to enrich the model with domain-specific content efficiently.

Applying Retrieval Augmented Generation (RAG) Technique

The session also delved into how the Retrieval Augmented Generation (RAG) technique could be integrated within the H2OGPT platform to enhance model responses. RAG marries the generative power of LLMs with the precision of retrieving pertinent information from a database or document collection. This method enables the model to incorporate additional context not initially included in its training data, enriching the responses with greater accuracy and contextual depth.

Fine-Tuning LLMs

The process of fine-tuning LLMs was demystified, highlighting its accessibility via the H2O LLM Studio. This tool, part of the H2O suite, exemplifies how advanced machine learning techniques can be applied practically, demonstrating the ease with which models can be adjusted and enhanced. This session emphasized the tool's capacity to streamline the fine-tuning process, making it more approachable for users without extensive technical expertise in machine learning.

Open Discussion

The AICoP sessions included interactive discussions on the practical challenges of employing LLMs, touching upon data privacy, cost efficiency, computing and technical resources required. The dialogue highlighted the importance of balancing data privacy and ethical LLM deployment with cost-effective strategies for academic and research endeavors. 

Emerging Technologies' mission is to empower faculty, researchers, and administrative members to utilize next-generation technologies to explore, adopt and integrate new applications and strategies. Please visit our AI: Community of Practice for discussion highlights on AI topics and an interest intake form to join the monthly gatherings.