Large language models (LLMs) are often trained with instruction fine-tuning to adapt to downstream tasks to enhance their generalization ability, but this method has certain performance limitations for LLMs' classification tasks, and sometimes cannot meet the task requirements. To address this issue, a global feature extraction classification large model framework is proposed. This framework uses the global feature extraction enhancement method proposed in this paper to release global features in the attention layer, then enhance the features, and apply the depth low-rank fine-tuning optimization loss proposed in this paper during fine-tuning. Finally, a global feature extraction classification large model is constructed. Compared with the baseline model RoBERTa, the accuracy on the general sentiment analysis dataset SST-2 and AGNews was improved by 1.44 and 0.95 percentage points, respectively. Compared with the baseline model PIQN, the F₁ score on the general named entity recognition (NER) dataset OntoNotes and CoNLL2003 was improved by 0.79% and 1.99%, respectively. The experimental results show that, under the condition of not requiring complex prompt engineering or external knowledge, the performance of the large model using this framework is significantly better than that of its several times larger LLMs.