Are Human Conversations Special? A Language Model Perspective

Large language models (LLMs) have shown remarkable capabilities across various tasks but their performance in handling human-human interactions remains a question. Human conversations are complex and multifaceted, with interactivity, contextuality, adaptability, and emotional states as key characteristics. However, an analysis of LLM training data reveals that human conversation data is underrepresented, constituting only 0.0085% of the total data. This imbalance leads to artifacts that hinder a model's ability to effectively handle such underrepresented domains. Quantitative and qualitative analyses show that human conversations demand significantly longer attention distances compared to other data types, indicating the need for more robust modeling of long-term contextual relationships by models. To address this issue, Symbl has developed Nebula, an LLM specialized for human conversations, trained with a significant amount of conversation data and optimized to work best on such interactions.