Abstract: Abstract Aim This study aims to develop a methodological framework for plant trait extraction using large language models (LLMs) and systematically evaluate the effects of model type, prompt design, and parameter settings on task performance. Methods Using the Flora of China as the data source, we constructed a standardized benchmark dataset containing 27,922 manually annotated records and designed a unified data preprocessing pipeline and systematic prompt templates. Two models (DeepSeek-V3 and DeepSeek-R1-0528-Qwen3-8B) were selected, and a three-factor experimental design was established by combining model type, temperature setting, and prompt strategy. Multiple repeated experiments were conducted for plant growth form and life span classification tasks. Model performance was comprehensively evaluated using missing rate, mismatch rate, accuracy, precision, recall, specificity, F1-score, and AUC, and analysis of variance was used to test the significance of main effects and interaction effects. Important findings The results showed that different models varied in parameter sensitivity and error-type control. DeepSeek-V3 showed more stable overall performance, whereas DeepSeek-R1-0528-Qwen3-8B was more sensitive to systematic prompts. A lower temperature setting generally helped reduce hallucination rates and the risk of output instability, whereas a higher temperature setting (0.6) may improve recall in some scenarios. Systematic prompts significantly reduced missing and hallucination rates, improved output completeness and category validity, and enhanced classification performance in some tasks and metrics. Compared with simple model performance comparison, parameter configuration and prompt strategy had a more critical influence on the practical applicability of LLMs in botanical text parsing tasks. The proposed framework provides a reusable technical pathway for automated biodiversity information extraction.

Key words: large language models, plant functional traits, text mining, biodiversity informatics, evaluation framework construction