Fine-Tuning General-Purpose Large Language Models for Agricultural Applications:A Reproducible Framework and Evaluation Protocol Based on Qwen3-8B

# Fine-Tuning General-Purpose Large Language Models for Agricultural Applications: A Reproducible Framework and Evaluation Protocol Based on Qwen3-8B
Source: [https://arxiv.org/html/2606.28992](https://arxiv.org/html/2606.28992)
Zhaoyang Li1,\*Ruijie Zhang1Jiaqi Liu2Zhaoji Sun1 1Sanya University 2Hebei International Studies University \*Corresponding author: Zhaoyang Li,[beijizhi@outlook\.com](https://arxiv.org/html/2606.28992v1/mailto:[email protected])

###### Abstract

General\-purpose large language models \(LLMs\) have demonstrated strong abilities in open\-domain question answering, information extraction, and text generation\. Agricultural applications, however, are domain\-specific, region\-dependent, time\-sensitive, and safety\-critical\. Without data governance, expert evaluation, and evidence constraints, an agricultural assistant may produce unreliable advice on crop diseases, pesticide use, fertilization, or policy interpretation\. This paper does not report any model\-performance claims that have not been produced by an actual training run and expert evaluation\. Instead, we propose AgriTune\-R, a reproducible and auditable framework for adapting general\-purpose LLMs to agricultural tasks\. The framework selects the publicly verifiable Qwen3\-8B model as the recommended base model and integrates agricultural data governance, instruction construction, LoRA/QLoRA parameter\-efficient fine\-tuning, retrieval\-augmented generation, expert evaluation, and safety control for high\-risk questions\. The contributions are: \(1\) a structured workflow for agricultural LLM adaptation; \(2\) an evaluation protocol for agricultural knowledge QA, pest and disease consultation, cultivation management, and policy explanation; \(3\) an expert\-review rubric combining factuality, safety, evidence consistency, and uncertainty expression; and \(4\) a clear separation between protocol design and empirical conclusions, providing an executable baseline for future empirical studies\.

Keywords:agricultural LLMs; Qwen3\-8B; domain adaptation; LoRA; QLoRA; retrieval\-augmented generation; agricultural question answering; model evaluation

## 1Introduction

Agricultural decision\-making relies on crop physiology, disease symptoms, soil conditions, weather, local policy, pesticide registration, and practical cultivation experience\. Although general\-purpose LLMs can process natural\-language questions, their parametric knowledge is not guaranteed to contain current, local, or safety\-compliant agricultural information\. In farmer\-facing consultation, a fluent but incorrect answer may lead to pesticide misuse, delayed disease control, economic loss, or food\-safety risks\. Consequently, agricultural LLM research should evaluate not only fluency but also data provenance, evidence chains, expert review, risk control, and deployment responsibility\.

Recent studies suggest that agricultural foundation models and domain\-specific LLM systems are promising, but they also face challenges related to data quality, benchmark design, hallucination, and deployment safety\[[8](https://arxiv.org/html/2606.28992#bib.bib8),[9](https://arxiv.org/html/2606.28992#bib.bib9),[10](https://arxiv.org/html/2606.28992#bib.bib10),[11](https://arxiv.org/html/2606.28992#bib.bib11)\]\. AgriBERT injects agricultural and food\-related knowledge into language models for matching food descriptions and nutrition data\[[7](https://arxiv.org/html/2606.28992#bib.bib7)\]\. AgroGPT focuses on agricultural vision\-language instruction construction and multimodal dialogue\[[9](https://arxiv.org/html/2606.28992#bib.bib9)\]\. ShizishanGPT integrates retrieval\-augmented generation, knowledge graphs, and tool use for agricultural question answering\[[10](https://arxiv.org/html/2606.28992#bib.bib10)\]\. AgriGPT further emphasizes data engines, domain\-specific benchmarks, and multi\-channel retrieval augmentation\[[11](https://arxiv.org/html/2606.28992#bib.bib11)\]\. These studies show that an agricultural assistant should be assessed not merely by whether it answers, but by whether its answer is grounded, safe, and reviewable\.

This paper clarifies its research position\. It is not a report claiming that a model has already been trained and has achieved numerical improvements\. Instead, it is a methods and evaluation\-protocol paper intended to provide a complete, executable, and auditable foundation for future empirical fine\-tuning studies\. To ensure model\-source accuracy and reproducibility, we select Qwen3\-8B as the recommended base model\. The Qwen3 technical report and official repository list publicly available Qwen3 model scales including 0\.6B, 1\.7B, 4B, 8B, 14B, 32B, and MoE variants; therefore, the 8B\-scale model is appropriate for a reproducible paper\[[1](https://arxiv.org/html/2606.28992#bib.bib1),[2](https://arxiv.org/html/2606.28992#bib.bib2)\]\.

The main contributions of this paper are as follows:

1. 1\.We propose AgriTune\-R, a framework that integrates agricultural data governance, instruction fine\-tuning, retrieval\-augmented generation, and safety evaluation\.
2. 2\.We provide a Qwen3\-8B\-based adaptation protocol while avoiding unverifiable model names and unexecuted performance numbers\.
3. 3\.We design an agricultural evaluation protocol covering knowledge QA, pest and disease consultation, cultivation management, policy explanation, and high\-risk refusal\.
4. 4\.We present practical tables and execution rules for expert scoring, evidence consistency, risk control, and reproducibility logging\.

## 2Related Work

### 2\.1General LLMs and parameter\-efficient adaptation

General LLMs acquire language understanding, reasoning, and generation abilities through large\-scale pretraining\. Full fine\-tuning can be expensive and data\-demanding when such models are adapted to specialized domains\. LoRA freezes pretrained weights and injects trainable low\-rank matrices, reducing the number of trainable parameters\[[4](https://arxiv.org/html/2606.28992#bib.bib4)\]\. QLoRA trains LoRA adapters through quantized base\-model weights, further reducing memory requirements\[[5](https://arxiv.org/html/2606.28992#bib.bib5)\]\. These methods make agricultural adaptation feasible for universities, local extension services, and small research teams with limited computing resources\.

### 2\.2Retrieval\-augmented generation

Agricultural knowledge changes over time\. Pesticide registration, local policy, pest and disease alerts, and weather\-related recommendations may become outdated\. Relying only on model parameters can therefore produce obsolete or unsupported answers\. Retrieval\-augmented generation \(RAG\) supplies external evidence to the generation model, making answers more traceable\[[6](https://arxiv.org/html/2606.28992#bib.bib6)\]\. In agriculture, RAG is particularly useful for extension manuals, policy explanation, disease\-control guidance, and safety\-standard checking\.

### 2\.3Agricultural language and multimodal models

Agricultural model research is moving from single\-task models to domain foundation models and model ecosystems\. AgriBERT addresses semantic matching in agricultural and food\-related text\[[7](https://arxiv.org/html/2606.28992#bib.bib7)\]\. Reviews of smart\-agriculture foundation models emphasize the need to connect language, vision, multimodal, and decision\-making tasks while accounting for data quality and deployment risk\[[8](https://arxiv.org/html/2606.28992#bib.bib8)\]\. AgroGPT shows that expert\-tuned agricultural vision\-language data can reduce domain gaps\[[9](https://arxiv.org/html/2606.28992#bib.bib9)\]\. ShizishanGPT and AgriGPT highlight the importance of retrieval, knowledge graphs, tool use, and domain\-specific evaluation in agricultural QA systems\[[10](https://arxiv.org/html/2606.28992#bib.bib10),[11](https://arxiv.org/html/2606.28992#bib.bib11)\]\.

## 3Model Selection and Truthfulness Principles

Truthfulness is a prerequisite for the proposed study design\. Any content that is not supported by an actual training run, inference experiment, or expert evaluation is not written as an empirical conclusion\. Model selection follows three principles: public verifiability, community reproducibility, and feasible computing cost\. Based on these principles, we recommend Qwen3\-8B rather than an unverifiable model name\.

Table 1:Principles for selecting the base model
## 4The AgriTune\-R Framework

AgriTune\-R decomposes agricultural adaptation into auditable steps, addressing common problems such as unclear data provenance, opaque training, and insufficient evaluation\. The full workflow is shown in Figure[1](https://arxiv.org/html/2606.28992#S4.F1)\.

Agricultural datasource, license, timelinessInstruction datatask, evidence, riskQwen3\-8Bverifiable base modelLoRA/QLoRAPEFT adaptationExpert review and safety auditfactuality, evidence, safetyKnowledge basepolicy/standards/manualsRAGevidence in contextThe figure is a workflow description only; no performance numbers are reported before real training and expert evaluation\.

Figure 1:Overview of AgriTune\-R\. The figure describes the method workflow and does not report experimental results\.### 4\.1Data governance

Agricultural data should be drawn from authoritative, licensed, and traceable sources, such as government documents, extension manuals, crop\-cultivation textbooks, pesticide labels and registration rules, agricultural standards, expert\-reviewed QA, and properly licensed datasets\. Each sample must keep source, publication date, license status, task type, evidence passage, and review record\. The governance pipeline is shown in Figure[2](https://arxiv.org/html/2606.28992#S4.F2)\.

1\. Source recordinstitution, date, region2\. License checkcopyright and scope3\. Quality filteringdeduplication, privacy, time4\. Expert reviewterms and safety boundaries5\. Version releasedata card and audit logTraceable sourceLicensed usePrivacy removedReviewableVersioned releaseEach sample stores source, date, license status, task type, evidence passage, and review record\.

Figure 2:Agricultural data governance and sample construction pipeline\.Table 2:Admission and exclusion rules for agricultural data sources
### 4\.2Task and instruction construction

Agricultural instruction data should contain not only a question and an answer, but also task type, crop, region, growth stage, evidence, risk level, and uncertainty notes\. A recommended sample structure is:

> instruction: Explain the typical symptoms and control principles of rice blast based on the evidence\. context: Crop=rice; region=South China; growth stage=tillering; evidence=Section X of an extension manual\. response: Describe symptoms, possible causes, field\-confirmation steps, control principles, and cases requiring local expert consultation\. safety: Do not provide off\-label pesticide dosages; do not replace an in\-person plant\-protection diagnosis\.

### 4\.3Parameter\-efficient fine\-tuning plan

Let the pretrained weights beW0W\_\{0\}\. LoRA represents a weight update as a low\-rank decomposition:

W=W0\+Δ​W=W0\+B​A,W=W\_\{0\}\+\\Delta W=W\_\{0\}\+BA,\(1\)whereA∈ℝr×dA\\in\\mathbb\{R\}^\{r\\times d\},B∈ℝd×rB\\in\\mathbb\{R\}^\{d\\times r\}, andrris the rank\. During training,W0W\_\{0\}is frozen and only the low\-rank matrices are optimized\. If QLoRA is used, the base model is stored in quantized form while gradients are propagated through LoRA adapters\. We recommend Qwen3\-8B as the base model for future empirical experiments and require all training configurations to be disclosed in the model card\.

Figure[3](https://arxiv.org/html/2606.28992#S4.F3)shows the recommended LoRA/QLoRA adaptation structure and clarifies which modules should be recorded as trainable or frozen in a future empirical run\.

Agricultural instructionevidence and risk tagsContext packinginput and targetFrozen baseQwen3\-8BSupervised lossadapter updateAgricultural adapterversioned releaseTrainable LoRAAttentionTrainable LoRAMLP/FFN

Figure 3:LoRA/QLoRA adaptation structure based on Qwen3\-8B\. Gray modules denote trainable adapters; the figure describes the training structure and does not claim completed training\.Table 3:Recommended configuration records for reproducible empirical fine\-tuning
### 4\.4Retrieval\-augmented generation

The agricultural knowledge base should be maintained as an independent module\. Each entry should include the original text, source, publication date, applicable region, topic tags, and citable passages\. For high\-risk agricultural questions, the model should answer based on retrieved evidence\. If evidence is insufficient, the model should express uncertainty or recommend consultation with local agricultural experts\. RAG is not a guarantee of correctness: retrieval recall, evidence ranking, and citation quality must be evaluated separately\.

Figure[4](https://arxiv.org/html/2606.28992#S4.F4)further specifies the evidence flow of agricultural RAG\. Source filtering and evidence\-sufficiency checks should be completed before generation\.

User questioncrop/region/symptomsQuery rewritingterm normalizationAgricultural KBpolicy/standards/manualsHybrid retrievalkeyword \+ vectorRerank and filtertime/region/licenseEvidence promptpassages \+ limitsAnswer generationbasis/advice/uncertaintydowngrade or refuse if evidence is insufficient

Figure 4:Agricultural retrieval\-augmented generation pipeline\. The figure emphasizes evidence selection and uncertainty handling rather than performance\.

## 5Evaluation Protocol

Agricultural LLM evaluation should not rely only on generic QA metrics\. This paper divides evaluation into five tasks: agricultural knowledge QA, pest and disease consultation, cultivation management, policy explanation, and high\-risk refusal\. Each task requires a different set of dimensions, as shown in Table[4](https://arxiv.org/html/2606.28992#S5.T4)\.

After a real experiment is completed, evaluation should not stop at automatic metrics\. Figure[5](https://arxiv.org/html/2606.28992#S5.F5)gives a closed loop for expert review, disagreement adjudication, and data improvement\.

Stratified test settask/crop/region/riskModel outputanswer \+ evidenceAutomatic checkformat/citation/forbidden itemsDual expert reviewfactuality/safety/actionabilityAdjudicationerror taxonomyEvaluation cardmethod, limits, examplesfeed errors back to data governance

Figure 5:Agricultural model\-evaluation loop for human review and error analysis after real experiments\.Table 4:Coverage of tasks and required evaluation dimensions\. The symbol✓\\checkmarkindicates required evaluation, not a model score\.Table 5:Expert\-review rubricTable 6:Recommended composition of the test set
## 6Safety Control and Responsibility Boundaries

High\-risk agricultural questions include pesticide dosage, disease misdiagnosis, food safety, policy eligibility, extreme\-weather response, and economic input decisions\. A model must not replace the final judgment of local extension experts, plant\-protection personnel, or government authorities\. High\-risk questions should follow risk detection, evidence retrieval, evidence\-sufficiency judgment, answer generation, refusal or human escalation, and audit logging, as shown in Figure[6](https://arxiv.org/html/2606.28992#S6.F6)\.

User questioncrop, region, symptomsRisk detectionpesticide, diagnosis, policyEvidence retrievalstandards, labels, manualsSufficient evidence?Evidence\-grounded answerwith boundariesRefuse or escalatesafer alternativeAudit logevidence, risk, outputyesno

Figure 6:Safety\-control workflow for high\-risk agricultural questions\.Table 7:Handling rules for high\-risk agricultural questionsFrom a system\-deployment perspective, an agricultural QA system should include input anonymization, risk classification, retrieval service, output guardrails, and human escalation\. Figure[7](https://arxiv.org/html/2606.28992#S6.F7)shows these responsibility boundaries\.

Farmer/extension usermobile or webInput gatewayanonymization and regionRisk classifierordinary/high\-riskRAG serviceevidence retrievalHuman escalationexpert or local authorityQwen3\-8B \+ adaptercontrolled generationOutput guardrailsafety and citation checkAuditable answeradvice/evidence/limitslogging and review

Figure 7:Deployment and responsibility boundaries for an agricultural QA system\. The diagram emphasizes auditability, privacy, and human escalation\.
## 7Reproducibility Plan

To make future empirical studies reproducible, researchers should release or submit the following materials with their model or paper:

1. 1\.A data card describing data source, license, time range, region coverage, anonymization, and exclusion rules\.
2. 2\.A model card describing base model, adaptation method, training configuration, risk boundary, and known limitations\.
3. 3\.An evaluation card describing test\-set construction, number of experts, inter\-rater agreement, evidence consistency, and error categories\.
4. 4\.An audit log recording high\-risk questions, refusal examples, insufficient\-evidence cases, and human\-review outcomes\.

Table 8:Truthfulness checklist before submission or release
## 8Limitations

The limitation of this paper is explicit: it does not execute real Qwen3\-8B agricultural fine\-tuning and does not report model\-performance results\. Therefore, it cannot prove that one fine\-tuning configuration is superior to another\. Its value lies in providing a truthful methods framework, data\-governance pipeline, and evaluation protocol\. Future work should add experimental results based on real agricultural data, real training logs, and expert evaluation, including failure cases and safety boundaries\.

## 9Conclusion

This paper proposes AgriTune\-R, a reproducible framework for adapting LLMs to agricultural applications\. In contrast to writing simulated data as if it were a complete empirical result, this work follows a truthfulness principle: no numerical performance improvement is reported before real training and expert evaluation are completed\. We select publicly verifiable Qwen3\-8B as the recommended base model and provide data governance, LoRA/QLoRA adaptation, retrieval\-augmented generation, expert scoring, and high\-risk safety\-control procedures\. The paper can serve as a protocol foundation for future empirical agricultural LLM studies and for building reliable, safe, and auditable agricultural QA systems\.

## References

- \[1\]An Yang et al\. Qwen3 Technical Report\. arXiv:2505\.09388, 2025\.
- \[2\]QwenLM\. Qwen3 official repository\.[https://github\.com/QwenLM/Qwen3](https://github.com/QwenLM/Qwen3)\.
- \[3\]Qwen Team\. Qwen3\-8B model card\.[https://huggingface\.co/Qwen/Qwen3\-8B](https://huggingface.co/Qwen/Qwen3-8B)\.
- \[4\]Edward J\. Hu et al\. LoRA: Low\-Rank Adaptation of Large Language Models\. arXiv:2106\.09685, 2021\.
- \[5\]Tim Dettmers et al\. QLoRA: Efficient Finetuning of Quantized LLMs\. arXiv:2305\.14314, 2023\.
- \[6\]Patrick Lewis et al\. Retrieval\-Augmented Generation for Knowledge\-Intensive NLP Tasks\. NeurIPS, 2020\.
- \[7\]Saed Rezayi et al\. AgriBERT: Knowledge\-Infused Agricultural Language Models for Matching Food and Nutrition\. IJCAI, 2022\.
- \[8\]Jiajia Li et al\. Large Language Models and Foundation Models in Smart Agriculture: Basics, Opportunities, and Challenges\. arXiv:2308\.06668, 2023\.
- \[9\]Muhammad Awais et al\. AgroGPT: Efficient Agricultural Vision\-Language Model with Expert Tuning\. arXiv:2410\.08405, 2024\.
- \[10\]Shuting Yang, Zehui Liu, and Wolfgang Mayer\. ShizishanGPT: An Agricultural Large Language Model Integrating Tools and Resources\. arXiv:2409\.13537, 2024\.
- \[11\]Bo Yang et al\. AgriGPT: a Large Language Model Ecosystem for Agriculture\. arXiv:2508\.08632, 2025\.