An Emotion Text Classification Model Based on Llama3-8b Using Lora Technique
Hongyi Shui, Yuanjing Zhu, Fan Zhuo, Yibo Sun, Daoyuan Li
Abstract
This study introduces a method utilizing the Llama3-8b model for emotion text classification. The training process is accelerated by incorporating Lora and FlashAttention techniques. On an emotion text dataset containing six categories, our improved Llama3-8b model demonstrates superior performance in emotion classification, surpassing other transformer models including Bert-Base, Bert-Large, Roberta-Base, and Roberta-Large. The Llama3-8b model, fine-tuned with supervised learning, achieved an accuracy of 0.9262, outperforming all other models and highlighting the advantages of large language models in classification tasks. This study illustrates the potential of specialized fine-tuning techniques in enhancing the performance of language models in domain-specific tasks, particularly in emotion text classification.