This paper proposes Baichuan2-Sum, an instruction fine-tuned model for role-oriented dialogue summarization. The model is based on the Baichuan2 large language model and is trained on two public dialogue summarization datasets: CSDS and SAMSUM. The model is designed to generate summaries for different roles in a dialogue by setting different instructions for each role. To improve performance, the NEFTune technique is applied, which adds noise to the embedding layer during training. The experiments show that Baichuan2-Sum achieves state-of-the-art results on both datasets, with significant improvements in ROUGE scores compared to previous models. On the SAMSUM dataset, the model shows a 21% increase in ROUGE-1, a 32% increase in ROUGE-2, and a 9% increase in ROUGE-L. The model is publicly available on GitHub, and the code supports training and evaluation for other large language models like LLaMA2, Bloom, and ChatGLM. The paper also discusses related works, including instruction fine-tuning and noisy embedding techniques, and presents the model's architecture, training details, and evaluation results. The model demonstrates superior performance in terms of accuracy, coherence, and grammatical correctness in human evaluations. The model is trained on a single RTX 4090 GPU in less than 6 hours for CSDS and 4 hours for SAMSUM. The results show that the model is effective for dialogue summarization tasks and can be used for future research.This paper proposes Baichuan2-Sum, an instruction fine-tuned model for role-oriented dialogue summarization. The model is based on the Baichuan2 large language model and is trained on two public dialogue summarization datasets: CSDS and SAMSUM. The model is designed to generate summaries for different roles in a dialogue by setting different instructions for each role. To improve performance, the NEFTune technique is applied, which adds noise to the embedding layer during training. The experiments show that Baichuan2-Sum achieves state-of-the-art results on both datasets, with significant improvements in ROUGE scores compared to previous models. On the SAMSUM dataset, the model shows a 21% increase in ROUGE-1, a 32% increase in ROUGE-2, and a 9% increase in ROUGE-L. The model is publicly available on GitHub, and the code supports training and evaluation for other large language models like LLaMA2, Bloom, and ChatGLM. The paper also discusses related works, including instruction fine-tuning and noisy embedding techniques, and presents the model's architecture, training details, and evaluation results. The model demonstrates superior performance in terms of accuracy, coherence, and grammatical correctness in human evaluations. The model is trained on a single RTX 4090 GPU in less than 6 hours for CSDS and 4 hours for SAMSUM. The results show that the model is effective for dialogue summarization tasks and can be used for future research.