This model is an experimental fine-tuning model used for troubleshooting mobile network voice calls (VoLTE/VoNR). The input is a textual description of the call process, and the output includes problem localization, diagnostic conclusions, and root cause analysis. The dataset used for fine-tuning training comes from nearly 2,000 live-network cases. After manually analyzing the raw signaling in PCAP format, descriptions of signaling analysis and problem localization are provided. Based on these original materials, the training dataset is formatted. Since the number of cases is not large enough (not yet in the tens of thousands), the trainer is also a beginner with limited understanding of model training. The hardware configuration used is relatively low (a GPU with 24GB of memory), and only an 8B model can be outputted (although a 14B model was trained, weight merging could not be performed). Therefore, the model’s capability is not ideal. It mainly verifies the feasibility of the entire process operation. It is hoped that in future versions, there will be opportunities to continuously improve and enhance the model’s capability.

该模型是一个实验性的微调模型，用于对移动网语音呼叫 (VoLTE/VoNR) 做出故障诊断，输入为呼叫过程的文字描述，输出为问题定位、诊断结论和根因分析。 微调训练采用的数据集来自近2000例现网真实案例，人工分析 PCAP 格式的原始信令后给出信令分析和问题定位的描述，基于这部分原始资料格式化成训练数据集。 由于案例数并不够多（还没上万），训练者也是初学者，对模型训练的理解有限，使用的硬件配置也比较低（24GB 显存的GPU），只能输出8B模型（训练了14B模型却无法执行权重合并），因此模型能力并不理想，主要还是验证了整个流程操作的可行性。希望在后续的版本中，还有机会不断改进和提高这个模型的能力。