Developing large language models for display industrial knowledge: Data augmentation, training techniques, and evaluation strategies

Abstract

Large Language Models (LLMs) can be applied to many fields in the display industry. However, general LLMs lack domain-specific knowledge and specialized terminology understanding, which results in inaccurate responses when applied to industrial question-answering(Q&A) scenarios. To address this issue, this work introduces a framework of Large Language Model training to effectively import the Display Industry Knowledge. This framework is specifically designed to enhance the comprehension ability of LLMs on the knowledge from the display industry field by improving specialized data governance, knowledge distillation techniques, data augmentation strategies, and continual pre-training mechanisms. This approach not only significantly improves the model's performance in Q&A applications within the display industry but also prevents catastrophic forgetting of common knowledge. Experimental results demonstrate the effectiveness of these techniques. We hope that this work can be also helpful for the customization of LLMs in other specialized domains.