Zero-Shot Instruction Generation via Dual-Alignment Instruction Wrappers with Summary-Text Fused Instruction Wrappers

With the deepening application of large language models in the field of natural language processing, ensuring the diversity, knowledgeability, and integrity of instruction data has become crucial for enhancing model performance and domain adaptation capabilities. Focusing on these aspects, this paper proposes a method for constructing instruction data by fusing document summaries and text fragments, which improves the… of the instruction dataset, and thus improve the performance of the model fine-tuned with the dataset. Diverse question-answer pairs covering multiple scenarios and dimensions are generated through ChatGPT-4, and multiple semantic alignment screening strategies are combined to ensure semantic consistency and integrity of the data, effectively reducing the risk of model hallucinations. Experiments conducted on mainstream models such as ChatGLM, Qwen, and DeepSeek, as well as datasets in general and medical domains, demonstrate that this method significantly improves the diversity and integrity of model instruction generation. The optimized models, in tasks such as medical question-answering, can generate more knowledgeable and logically coherent long-text responses compared to unfine-tuned models and the ChatGPT-3.5 baseline, providing an effective path with both theoretical depth and practical value for the instruction tuning and domain adaptation of large language models.