ChickVLM: Domain-Adaptive vision-language model for interactive chicken disease diagnosis

• We propose ChickVLM, a domain-adaptive vision-language model for poultry diagnosis. • A knowledge-guided pipeline integrates texts, images, and diagnostic dialogues. • A high-quality multimodal dataset (CMDD) with knowledge-driven diagnostic dialogues is constructed. • A three-stage training pipeline enhances domain alignment and interpretability. • ChickVLM outperforms the strongest baseline, Med-Flamingo, on both classification and dialogue generation tasks. Rapid and precise diagnosis of poultry diseases is critical for ensuring biosecurity and mitigating economic losses in large-scale production environments. Traditional diagnostic workflows often rely on scarce veterinary expertise, leading to delayed interventions and subjective assessments. While multimodal large language models (MLLMs) offer promising avenues for automated diagnosis, their practical application in agriculture is hindered by a lack of domain-specific multimodal data and specialized task alignment. In this study, we present ChickVLM, a specialized vision-language model designed for interactive and interpretable chicken disease diagnosis. To empower the model, we developed a knowledge-guided data construction pipeline that integrates expert-validated veterinary literature with real-world clinical pathological images, resulting in the Chicken Disease Multimodal Diagnostic Dataset (CMDD)—a high-fidelity resource comprising 9,000 image-text pairs of multi-turn diagnostic dialogues. Built upon the Qwen2.5VL-7B architecture, ChickVLM is refined through a three-stage training pipeline encompassing continued pre-training, concept alignment, and instruction fine-tuning. To ensure high parameter efficiency suitable for agricultural electronic instrumentation, LoRA-based fine-tuning was strategically employed. Experimental results demonstrate that ChickVLM consistently outperforms the state-of-the-art baseline, Med-Flamingo, in both diagnostic reasoning and disease classification. Specifically, ChickVLM achieves relative improvements of 5.56% in BLEU-4, 5.37% in ROUGE-L, 5.05% in accuracy, and 6.30% in F1 score. These findings underscore ChickVLM’s potential as a robust decision-support tool for intelligent avian health management at the farm level.