The development of Large Language Models (LLMs) such as GPT and LLaMA has significantly revolutionized natural language processing (NLP). They have found use in a broad range of functions, causing a growing demand for custom LLMs amongst individuals and corporations. However, the development of these LLMs is resource-intensive, posing a significant challenge for potential users.
To address this problem, experts have suggested knowledge fusion of LLMs as an alternate way to build robust models while minimizing development costs. This involves integrating multiple LLMs into a unified system to take advantage of their strengths across varying tasks. In the past, achieving such integration required ensemble methods or direct merging of neural networks, which often resulted in inefficiencies during analyses or necessitated uniform network designs for combination.
FUSELLM has introduced a pioneering paradigm for knowledge fusion circumventing these limitations. This employs probability distribution matrices produced by several LLMs to transfer collective knowledge into a target LLM through minimal continuous training. Through this technique, FUSELLM can fuse pre-trained LLMs of different designs and sizes into one cohesive model.
Based on FUSELLM principles, the study introduces FUSECHAT, a system specifically designed for the fusion of chat LLMs, irrespective of their design and scale. It integrates knowledge from different LLMs and merges the capability within the parameter space to incorporate the collective knowledge from the source models. FUSECHAT employs a novel technique called VARM (Variation Ratio Merge) to determine the weights for combination based on the variation ratio of parameter matrices before and after fine-tuning, permitting for detailed merging without the need for added training efforts.
An empirical evaluation using representative open-source chat LLMs showed FUSECHAT’s effectiveness at outperforming individual LLMs across various scales. The results from MT-Bench, an assessment benchmark for multi-turn dialogue capacity, showed FUSECHAT’s superior performance—made possible through the innovative VARM merging method. With its robust and scalable nature, FUSECHAT is a promising solution for integrating chat models within the fast-evolving open-source LLM development landscape.
FUSECHAT is a significant advancement in the LLM integration field, particularly for chat-based applications. Through knowledge fusion techniques, it provides a practical and efficient solution to consolidating the capabilities of various chat LLMs, addressing the challenge of resource-intensive model development. Its ability to integrate models of varying designs and scales and the VARM merging method’s effectiveness positions FUSECHAT as a versatile tool for enhancing dialogue system performance. As the demand for sophisticated AI chat systems continues to surge, FUSECHAT is set to play a pivotal role in prompting innovation and advancement in the sector.
