SeedBench: A Multi-task Benchmark for Evaluating Large Language Models in Seed Science

English | 简体中文

SeedBench is the first multi-task benchmark designed to evaluate large language models (LLMs) in seed science, focusing on seed breeding. This repository includes the dataset, evaluation code, and documentation to support research in this domain. Here is the usage.

SeedBench assesses LLMs across three core seed breeding stages:

• Gene Information Retrieval
• Gene Function and Regulation Analysis
• Variety Breeding with Agronomic Trait Optimization

Breeding Expert Workflow Framework

Built with domain experts, SeedBench features 2,264 expert-validated questions across 11 task types and 10 subcategories, initially targeting rice breeding. Future updates will include other crops like maize, soybean, and wheat.

• Corpus: 308,727 publications cleaned to 1.1 billion tokens; 279 segments from 113 documents.

• Questions: 2,264 across 11 task types, bilingual (English/Chinese), expert-validated.

• Focus: Rice breeding as a representative case.

  Types and metrics:

  Type ID	Question Type	Metric	Count
  Q&A
  QA-1	Multiple Choice	Accuracy	200
  QA-2	Multiple Answer	Macro-F1	187
  QA-3	Fill-in-the-Blank	ROUGE-L	224
  QA-4	Generation	ROUGE-L	242
  Summarization
  SUM-1	Simple Summarization	ROUGE-L	225
  SUM-2	Key Information Extraction	ROUGE-L	225
  Reading Comprehension
  RC-1	Multiple Choice	Accuracy	113
  RC-2	Multiple Answer	Macro-F1	108
  RC-3	Fill-in-the-Blank	ROUGE-L	221
  RC-4	Generation	ROUGE-L	240
  RC-5	Subcategory Classification	Accuracy	279

  Taxonomy Distribution:

  

We evaluated 26 LLMs, including proprietary, open-source, and domain-specific models. Highlights:

• Top Performers: DeepSeek-V3 (68.37), GPT-4 (67.88).

  

  

• Top Performers: DeepSeek-V3-671B (63.30), GPT-4 (62.06).

• base_model_eval/: Used to test base models without dialogue capabilities, i.e., evaluating performance after pretraining.
• sft_model_eval/: Used to test SFT (Supervised Fine-Tuning) models, with a total of 2,264 questions covering 10 subcategories (see Fig 2).
  • one-shot/: Organized by 11 task types (see Tab 1).
  • zero-shot/: Organized by 11 task types (see Tab 1).
• corpus/: 279 high-quality text segments and low-quality questions discarded after expert validation.
• README.md: This file.

To evaluate models on SeedBench, we utilize OpenCompass. Follow the steps below to set up the environment and run the evaluation.

Clone the OpenCompass repository and install the necessary dependencies (including modelscope for dataset downloading).

git clone https://github.com/open-compass/opencompass opencompass
cd opencompass
pip install -e .
pip install modelscope

Set the dataset source environment variable and execute the evaluation script. The example below uses Qwen/Qwen2.5-0.5B-Instruct.

DATASET_SOURCE=ModelScope python run.py --hf-type chat \
    --hf-path Qwen/Qwen2.5-0.5B-Instruct \
    --datasets seedbench_gen \
    --debug

📝 Notes:

• Dataset Download: The initial run may take a few minutes to automatically download the dataset from ModelScope.
• Local Models: You can replace Qwen/Qwen2.5-0.5B-Instruct with your absolute local path if necessary.
• Please see Here for details.

Open an issue on this repository for questions or contributions.

@inproceedings{ying-etal-2025-seedbench,
  title = "{S}eed{B}ench: A Multi-task Benchmark for Evaluating Large Language Models in Seed Science",
  author = "Ying, Jie  and
    Chen, Zihong  and
    Wang, Zhefan  and
    Jiang, Wanli  and
    Wang, Chenyang  and
    Yuan, Zhonghang  and
    Su, Haoyang  and
    Kong, Huanjun  and
    Yang, Fan  and
    Dong, Nanqing",
  editor = "Che, Wanxiang  and
    Nabende, Joyce  and
    Shutova, Ekaterina  and
    Pilehvar, Mohammad Taher",
  booktitle = "Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
  month = jul,
  year = "2025",
  address = "Vienna, Austria",
  publisher = "Association for Computational Linguistics",
  url = "https://aclanthology.org/2025.acl-long.1516/",
  pages = "31395--31449",
  ISBN = "979-8-89176-251-0",
  abstract = "Seed science is essential for modern agriculture, directly influencing crop yields and global food security. However, challenges such as interdisciplinary complexity and high costs with limited returns hinder progress, leading to a shortage of experts and insufficient technological support. While large language models (LLMs) have shown promise across various fields, their application in seed science remains limited due to the scarcity of digital resources, complex gene-trait relationships, and the lack of standardized benchmarks. To address this gap, we introduce SeedBench{---}the first multi-task benchmark specifically designed for seed science. Developed in collaboration with domain experts, SeedBench focuses on seed breeding and simulates key aspects of modern breeding processes. We conduct a comprehensive evaluation of 26 leading LLMs, encompassing proprietary, open-source, and domain-specific fine-tuned models. Our findings not only highlight the substantial gaps between the power of LLMs and the real-world seed science problems, but also make a foundational step for research on LLMs for seed design."
}