PaperlayerView exact revision on arXiv

Source-linked AI summary

AgriGPT: a Large Language Model Ecosystem for Agriculture

Bo Yang, Yu Zhang, Lanfei Feng, Yunkui Chen, Jianyu Zhang, Xiao Xu, Nueraili Aierken, Yurui Li, Yuxuan Chen, Guijun Yang, Yong He, Runhe Huang, Shijian Li

arXiv:2508.08632v1cs.AI

TL;DR

Agriculture remains underrepresented in AI research, despite decisions relying on diverse knowledge from crop science to market signals. AgriGPT combines a multi-agent data engine for agricultural instruction data, continual training, Tri-RAG retrieval, and the AgriBench-13K benchmark. AgriGPT retains strong performance on general-domain benchmarks and demonstrates effective multilingual transfer.

  • Problem

    Agriculture remains underrepresented in AI research, despite decisions relying on diverse knowledge from crop science to market signals.

    Source: Introduction · p. 1

  • Method

    AgriGPT combines a multi-agent data engine for agricultural instruction data, continual training, Tri-RAG retrieval, and the AgriBench-13K benchmark.

    Source: Introduction · p. 3, Data Engine · p. 3, Conclusion · p. 8

  • Results

    AgriGPT retains strong performance on general-domain benchmarks and demonstrates effective multilingual transfer.

    Source: Conclusion · p. 8

  • Takeaways & Limitations

    The model, data engine, dataset, and benchmark form a coherent ecosystem that advances agricultural AI research.

    Source: Conclusion · p. 8

    AgriGPT is text-only, relies on formal-source training data that lacks diversity, and does not explicitly handle regional dialects.

    Source: Potential Social Impact and Limitations · pp. 7–8

Abstract

from arXiv · show

Despite the rapid progress of Large Language Models (LLMs), their application in agriculture remains limited due to the lack of domain-specific models, curated datasets, and robust evaluation frameworks. To address these challenges, we propose AgriGPT, a domain-specialized LLM ecosystem for agricultural usage. At its core, we design a multi-agent scalable data engine that systematically compiles credible data sources into Agri-342K, a high-quality, standardized question-answer (QA) dataset. Trained on this dataset, AgriGPT supports a broad range of agricultural stakeholders, from practitioners to policy-makers. To enhance factual grounding, we employ Tri-RAG, a three-channel Retrieval-Augmented Generation framework combining dense retrieval, sparse retrieval, and multi-hop knowledge graph reasoning, thereby improving the LLM's reasoning reliability. For comprehensive evaluation, we introduce AgriBench-13K, a benchmark suite comprising 13 tasks with varying types and complexities. Experiments demonstrate that AgriGPT significantly outperforms general-purpose LLMs on both domain adaptation and reasoning. Beyond the model itself, AgriGPT represents a modular and extensible LLM ecosystem for agriculture, comprising structured data construction, retrieval-enhanced generation, and domain-specific evaluation. This work provides a generalizable framework for developing scientific and industry-specialized LLMs. All models, datasets, and code will be released to empower agricultural communities, especially in underserved regions, and to promote open, impactful research.

Introduction

Agriculture remains underrepresented in AI research, while agricultural LLM use is constrained by fragmented data, limited terminology coverage, weak retrieval grounding, and domain-specific reasoning gaps. AgriGPT addresses these needs through an open ecosystem combining scalable instruction curation, Tri-RAG, and a multi-task benchmark.

Introduction

  • Agricultural decisions draw on knowledge ranging from crop science to market signals, and the research gap is described as especially consequential for smallholder farmers facing knowledge constraints.The authors frame release of the code, data, and benchmarks as a way to lower agricultural AI deployment barriers.Source: Introduction · pp. 1–2
  • Agricultural LLMs face sparse terminology coverage, limited retrieval grounding, and domain-specific reasoning gaps, amid fragmented data and scarce experts.Deployment also requires local relevance, interpretability, and infrastructure awareness.Source: Introduction · p. 1
  • AgriGPT is presented as an open, unified ecosystem, contrasting with prior models described as lacking generative reasoning, relying on complex pipelines, or not being open-source.The authors also describe its engine and Tri-RAG as a transferable framework for other real-world fields.Source: Introduction · pp. 1, 3
  • AgriGPT combines a multi-agent engine for Agri-342K, Tri-RAG retrieval and reasoning, and AgriBench-13K for agricultural evaluation.The benchmark covers factual QA, diagnostic reasoning, and multi-hop reasoning.Source: Introduction · pp. 1–3
  • The multi-agent data engine curates 342K agricultural instruction examples and a multilingual version for cross-lingual capabilities.The ecosystem also includes Mini-AgriBench600 and a multilingual variant for lightweight and cross-lingual evaluation.Source: Introduction · pp. 2–3

AgriGPT

AgriGPT combines a multi-agent data engine, domain adaptation and supervised fine-tuning, Tri-RAG retrieval, and a dedicated agricultural benchmark. Its methods describe how the dataset and benchmark are built and how retrieval augments generation.

Data Engine

  • The data engine gathers credible documents across 9 agricultural domains, using over 600 sub-area keywords to guide retrieval, filtering, and categorization.Its pipelines process literature, reformat public QA data, and expand roughly 1K expert-written seed prompts into additional QA pairs.Source: Data Engine · pp. 3–4
  • Over 342K reasoning-augmented QA pairs are produced by a reusable data engine for domain-specific instruction datasets.The engine draws on credible agricultural documents, public QA sources, and expanded expert-written seed prompts, then applies collaborative agents for rewriting, validation, and evaluation.Source: Data Engine · pp. 3–4

Continual Pretraining Stage

  • Continual pretraining adapts Qwen3-8B to agricultural terminology and literature while aiming to preserve general-language capabilities.The authors use LoRA-based training before fine-tuning on Agri-342K.Source: Continual Pretraining Stage · p. 4

Supervised Fine-Tuning Stage

  • Supervised fine-tuning on Agri-342K trains AgriGPT to answer agricultural questions in the dataset’s QA format using cross-entropy loss.The method updates LoRA adapter weights and optionally the base model’s top layers.Source: Supervised Fine-Tuning Stage · p. 4

Tri-RAG Stage

  • Tri-RAG merges dense semantic retrieval, BM25 sparse retrieval, and knowledge-graph reasoning, then reranks evidence for augmented generation.The graph is built from approximately 2 million factual triples; the paper reports improved performance on complex, knowledge-intensive agricultural tasks.Source: Tri-RAG Stage · pp. 4–5

AgriBench-13K

  • AgriBench-13K is designed to evaluate agricultural language understanding and reasoning across representative task types.Table 2 presents the task distribution; the supplied passages do not provide its individual task counts.Source: AgriBench-13K · p. 5, Table 2 · p. 5
  • AgriBench-13K contains 12,780 reviewed samples across 13 task types, with similarity filtering used to reduce training-data leakage.The benchmark is built from expert-defined task categories and 585 seed prompts; Mini-AgriBench600 is offered for efficient evaluation.Source: AgriBench-13K · p. 5

Result

AgriGPT performs strongly across agricultural tasks and general-domain benchmarks, with multilingual transfer and complementary gains from domain-specific training and RAG. The section also reports deployment potential and identifies limits in modality, training-data diversity, and dialect handling.

Comparative Experiments

  • AgriGPT leads on both automatic and LLM-based evaluation, scoring highest across all LLM-evaluated dimensions and outperforming many larger closed-source models.It performs especially well on multi-hop reasoning and sequential understanding across the majority of AgriBench-13K tasks.Source: Comparative Experiments · p. 7

Generalization Evaluation

  • AgriGPT performs comparably to Qwen3-8B on MMLU, ARC, and OpenBookQA, with differences within 0.5 points across datasets.The authors interpret this as retaining general reasoning and language skills after domain adaptation.Source: Generalization Evaluation · p. 7

Multilingual Evaluation

  • AgriGPT achieves reasonable BLEU and Meteor scores on Chinese and Japanese, indicating transfer of instruction-following ability across languages.The evaluation uses a multilingual version of the 600-example Mini-AgriBench600 subset.Source: Multilingual Evaluation · p. 7

Ablation Study

  • Domain-specific training and RAG each improve BLEU and Meteor, while their combination yields the best performance.The ablation describes the components as complementary.Source: Ablation Study · p. 7

Potential Social Impact and Limitations

  • AgriGPT reaches 44.15 token/s on a single RTX 4090 GPU, which the authors cite as supporting cost-effective edge deployment and scalable use in low-resource settings.The authors describe potential benefits for agricultural productivity, decision-making, and food security.Source: Potential Social Impact and Limitations · p. 7
  • AgriGPT currently supports text input only, relies on formal sources that limit training-data diversity, and does not explicitly handle regional dialects.Future work includes image and sensor inputs, informal texts and farmer dialogues, and broader dialect coverage.Source: Potential Social Impact and Limitations · pp. 7–8

Conclusion

The conclusion presents AgriGPT as an agricultural LLM ecosystem built from a multi-agent data engine, Agri-342K, Tri-RAG, and AgriBench-13K. It reports strong general-domain performance and multilingual transfer, and frames the ecosystem as supporting practical deployment in low-resource farming communities.

Conclusion

  • The paper combines a multi-agent data engine, Agri-342K, Tri-RAG, and AgriBench-13K into an ecosystem for complex agricultural tasks.The benchmark covers agricultural tasks and difficulty levels.Source: Conclusion · p. 8
  • AgriGPT retains strong performance on general-domain benchmarks and demonstrates effective multilingual transfer beyond agriculture-specific contexts.Source: Conclusion · p. 8
  • The authors present the ecosystem as enabling equitable, practical deployment in real-world, low-resource farming communities.Source: Conclusion · p. 8

Related papers

Loading 2508.08632v1…