Join Donghao Huang at SuperAI Singapore 2025 as he delves into the intricacies of optimizing and benchmarking open-source large language models (LLMs). With over 25 years of experience in software development and 25 patents to his name, Donghao brings a wealth of knowledge and innovation to the field of AI and machine learning.
In this insightful presentation, Donghao addresses the common issue of text repetition in open-source models, a phenomenon that can greatly detract from user experience. He highlights the challenges of adjusting repetition parameters, showcasing how an imbalance can lead to incomplete or repetitive responses. To tackle this issue, Donghao introduces a groundbreaking metric called Repetition Aware Performance (RAP), alongside an efficient repetition detection algorithm. These innovations enable developers to balance text coherence and repetition effectively, preserving content quality while enhancing model performance.
Utilizing comprehensive benchmarks such as the WebQuestionsSP and Microsoft MARCO datasets, Donghao demonstrates the effects of meticulous tuning on natural language processing tasks. He further reveals the significance of prompt engineering, specifically the usage of chat templates, which dramatically reduce text repetition and improve overall model output.
Discover how these cutting-edge strategies can transform the capabilities of LLMs in various applications, from fintech to translation. Don't miss out on the opportunity to stay at the forefront of AI development.
If you find this talk insightful, consider liking the video, commenting with your thoughts or questions, and subscribing to our channel for more expert discussions from SuperAI Singapore 2025.