LoRA-based Domain Adaptation of Qwen-7B for Veterinary and Livestock Text Generation: A Lexical and Semantic Evaluation of Base and Fine-Tuned Outputs

Large language models (LLMs) have demonstrated strong capabilities in natural language processing; however, their effectiveness in specialized domains depends on their ability to represent domain-specific terminology, conceptual relations, and contextual knowledge. This study evaluates the effect of Low-Rank Adaptation (LoRA)-based domain adaptation on the text-generation behavior of Qwen-7B in veterinary and livestock sciences. The base Qwen-7B causal language model was adapted using a domain-specific corpus compiled from veterinary and livestock literature. The adaptation process was conducted within an instructionformatted causal language modeling framework, in which unlabeled domain texts were transformed into structured prompt–response representations. The base and LoRA-adapted models were evaluated on an independent 100-question domain-specific test set using a multidimensional protocol that combined lexical overlap metrics, Bilingual Evaluation Understudy (BLEU), Recall-Oriented Understudy for Gisting Evaluation (ROUGE), and Metric for Evaluation of Translation with Explicit ORdering (METEOR), with the semantic similarity metric Bidirectional Encoder Representations from Transformers Score (BERTScore).. The results showed that LoRA-based adaptation did not produce uniform improvement across all evaluation dimensions. Instead, the adapted model showed higher BERTScore-based semantic similarity while exhibiting reduced surface-level lexical overlap. Therefore, the observed effect should be interpreted as a metric-dependent reorientation of model output rather than as a general performance gain. Overall, this study provides descriptive evidence that parameter-efficient fine-tuning can modify the generation behavior of LLMs in veterinary and livestock text-generation tasks, while also highlighting the need for multidimensionally supported, contamination controlled, and expert-informed evaluation before such models can be considered reliable for clinical or decision support applications.