Objective To address the issues of hallucinations, contextual logical inconsistencies, and inability to run on low-resource devices when large language models perform named entity recognition (NER) in agriculture.

Method Using DeepSeek-671B as the teacher model, domain knowledge was transferred to student models with fewer parameters. The student models selected were low-parameter versions of DeepSeek, Qwen, and Llama (1.5 billion, 7 billion, and 14 billion parameters, respectively), which underwent distillation and counterfactual reasoning training. Model performance was experimentally validated on the CropDiseaseNer dataset, a specialized agricultural disease dataset.

Result By comparing the performance of a series of distilled student models, the results showed that DeepSeek-14B achieved an entity recognition F1 score of 89.60% while requiring only 2.08% of the parameters of the teacher model. Its performance significantly outperformed both the general-purpose large model GPT-mini-14B (F1 score: 57.64%) and the domain-adapted model GLiNER (F1 score: 82.96%) based on a general LLM. Further analysis revealed that the DeepSeek student model, sharing the same architecture, demonstrated superiority over models with different architectures in recognizing long-tail categories such as disease entities and pathogen genus names, owing to its parameter alignment advantage.

Conclusion This study validates the effectiveness of knowledge distillation in NER tasks within the agricultural domain, offering a novel solution for entity recognition technology in resource-constrained scenarios.