GBNP-2026:面向畜禽遗传、育种、营养与生产的大语言模型专业问答能力评测基准OA

[Objective]Livestock-related QA is terminology-heavy and often involves indicators,parameters,and trade-offs in breeding and production plans.This study evaluates LLM professional usability using subjective responses and explicit scoring rubrics.[Methods]We built a Chinese vertical benchmark,the Genetics,Breeding,Nutrition,and Production Benchmark(GBNP-2026),with 510 subjective items(363 short-answer and 147 essay)spanning the same four domains,each with a reference answer,checkable scoring points,and domain tags.Nine open-and closed-source models were tested under a unified zero-shot,context-free protocol,scored by coverage and quality,followed by error attribution on low-score cases.[Results]Mean scores ranged from 64.03 to 83.99 and coverage from 70.6%to 91.9%.Essay items scored higher than short-answer items on average;breeding was the weakest domain for most models.Among 628 low-score samples,missing knowledge(37.3%)and hallucinations(25.0%)dominated.Missing-knowledge rates differed across domains(x²=8.42,P=0.038),highest in breeding(47.1%).[Conclusion]Compared with multiple-choice accuracy alone,coverage-plus-quality scoring helps separate capability gaps in this setting;strictly numerical or proof-style tasks still need a dedicated sub-benchmark.GBNP-2026 and the pipeline support reproducible comparison,error diagnosis,and data planning for SFT/RAG.