Unveiling Imitation Learning: Exploring the Impact of Data Falsity to Large Language Model

arXiv:2404.09717v1 Announce Type: cross
Abstract: Many recent studies endeavor to improve open-source language models through imitation learning, and re-training on the synthetic instruction data from state-of-the-art proprietary models like ChatGPT and GPT-4. However, the innate nature of synthetic data inherently contains noisy data, giving rise to a substantial presence of low-quality data replete with erroneous responses, and flawed reasoning. Although we intuitively grasp the potential harm of noisy data, we lack a quantitative understanding of its impact. To this end, …

abstract art arxiv chatgpt cs.ai cs.cl cs.lg data endeavor giving gpt gpt-4 however imitation learning impact language language model language models large language large language model nature proprietary proprietary models state studies synthetic synthetic data through training type