How Far Have We Come with Generative Model Tuning? - Langcon 2024

How far have we come in tuning generative models?

Super-large language models and natural language processing

<Langcon2024> shared diverse experiences and insights from industry and academia surrounding Natural Language Processing (NLP) and its core, Large Language Models (LLMs) ! We share the lively presentations.

I recommend this to these people

Conference presentation topics

1) Create your own search using LLM

How do I create my own AI service (GPTs) that finds and answers the information I need? I'll share the process of creating my own service using LLM, including connecting to a server, storing data, and adding search functionality using Google APIs.

2) Recipe for turning a model who speaks English well into a model who speaks Korean well

Adding Korean to an English model can lead to performance limitations. Or, it can lead to the problem of only being good at Korean. To overcome these issues, I'll share a newly implemented model (KoMistLlama-Pro-9B). I'll also introduce the features and functions of EasyLM, a framework for LLM learning.

3) Agent operation that evolves with real users

Liner, a startup that started as a highlighter utility, has evolved into a user-friendly agent. This presentation will cover the lessons learned and expertise gained from operating Liner's agent-based product for over a year, as well as its journey toward building a data flywheel.

4) How is your AI copyright?

With the advent of the era of large-scale language models (LLMs), copyright issues are emerging as a hot topic. This presentation addresses copyright issues surrounding data collection, training, and model-generated output in LLM models. We will also explore the various licenses and commercial uses of these models.

5) How about accelerating large-scale vector operations using FPGA?

As advanced document processing techniques like RAG gain traction, the importance of Vector Database (VDB) is also growing. However, it still faces numerous technical challenges. Sionic AI, in collaboration with MetisX, is exploring ways to optimize vector computation through FPGA acceleration, and this presentation presents the technical blueprint for this effort.

Things to note before taking the course