Merging pre-trained LLMs with mergekit
One of the eMERGing techniques in the Large Language Model (LLM) space is model merging. The big idea is to combine two or more LLMs into a single model.
This simple idea has proven to be a powerful pattern - for creating merged LLMs that perform much better across multiple benchmarks. In simple terms, it’s a hack to improve performance.
From first principles, merging is a method for abstracting and leveraging what individual Language Models pay attention to and how they act. This is done simply by taking pre-trained LLMs from multiple models and combining them.
This works very well! Even today you can observe specific merging algorithms at the top of Hugging Face’s LLM Leaderboard.
In this event, we’ll discuss big ideas behind merging LLMs. We will cover how merging relates to fine-tuning and alignment techniques and what transformer architecture-specific considerations we should keep in mind while merging.
Of course, we’ll also learn about important merging algorithms available off-the-shelf today through mergekit, including:
Linear, a.k.a. the OG averaging approach
Spherical Linear Interpolation (SLERP)
TrIm, Elect Sign and Merge (TIES-Merging), which helps resolve interference
Drops And REscales (DARE), a.k.a. Language Models are Super Mario
Finally, we’ll show you exactly how to leverage mergekit to do a merge! We’ll test our custom LLM’s performance qualitatively and then quantitatively against specific benchmark test data.
You’ll learn:
About model merging and emerging best practices for prototyping
How to use the mergekit framework to build your high-performance LLMs
How to evaluate a merged LLM and how to avoid potential pitfalls
Who should attend the event?
GenAI enthusiasts interested in learning more about models at the top of the leaderboard
Aspiring AI Engineers looking to build and fine-tune complex LLM applications
AI Engineering leaders who want to understand how to leverage open-source models
Speakers
Charles Goddard is the creator of the open source toolkit MergeKit, which he recently brought to Arcee.ai while joining their team as a Senior Research Engineer leading a team doing model merging research. Prior to Arcee, Charles was a software engineer with a distinguished track record in the AR/VR and aerospace industries, including a long tenure at NASA’s Jet Propulsion Laboratory. Both Arcee and Charles hold a strong commitment to the open source community and have pledged to keep MergeKit in the public domain, helping contribute to pushing the boundaries of technology and human knowledge.
Dr. Greg” Loughnane is the Co-Founder & CEO of AI Makerspace, where he is an instructor for their AI Engineering Bootcamp. Since 2021 he has built and led industry-leading Machine Learning education programs. Previously, he worked as an AI product manager, a university professor teaching AI, an AI consultant and startup advisor, and an ML researcher. He loves trail running and is based in Dayton, Ohio.
Chris “The Wiz” Alexiuk is the Co-Founder & CTO at AI Makerspace, where he is an instructor for their AI Engineering Bootcamp. During the day, he is also a Developer Advocate at NVIDIA. Previously, he was a Founding Machine Learning Engineer, Data Scientist, and ML curriculum developer and instructor. He’s a YouTube content creator YouTube who’s motto is “Build, build, build!” He loves Dungeons & Dragons and is based in Toronto, Canada.
Follow AI Makerspace on LinkedIn and YouTube to stay updated about workshops, new courses, and corporate training opportunities.