The False Promise of Imitating Proprietary LLMs (arxiv.org)

Recent research critiques the trend of finetuning weaker open-source language models (LMs) on outputs from proprietary systems like ChatGPT. This method, aimed at cheaply augmenting model performance by imitating a stronger counterpart, was initially thought to enhance instruction-following capabilities. Despite early positive feedback from human raters, further assessments revealed a significant gap in performance, particularly on tasks not well-represented in the imitation data. The study highlighted that while imitation models mimic ChatGPT’s style effectively, they struggle with factual accuracy, indicating that human evaluations could overlook these critical shortcomings. The findings challenge the assumption that model imitation can readily bridge the capabilities gap between open and closed systems. The authors argue that simply imitating proprietary models does not yield substantial improvements in performance, instead suggesting that efforts should focus on developing superior base models for open-source LMs. This conclusion holds significant implications for the AI/ML community, as it underscores the need for innovative advancements in model architecture and training methodologies rather than relying on imitation strategies that ultimately fall short.