Research map / Large language models / Instruction and fine-tuning

Empowering and Assessing the Utility of Large Language Models in Crop Science

NeurIPS 2024 · Large language models › Instruction and fine-tuning · approach: Calibration and knowledge in LLM tuning · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces CROP, a crop-science instruction-tuning dataset (210K single-turn and 1,871 multi-turn dialogues on rice and corn) and a 5,045-question bilingual MCQ benchmark, showing fine-tuning open LLMs on it improves accuracy.

Problem

LLMs are not yet reliable crop science assistants because no comprehensive instruction-tuning dataset or broad, bilingual benchmark exists for the crop science domain.

Approach

A task-oriented, LLM-centric pipeline (GPT-4 generation from professional books and research papers, RAG-grounded multi-agent multi-turn dialogue via a CAMEL-style domain adaptation, and LLM-plus-human filtering) produces a large crop-specific instruction dataset, and a difficulty-stratified MCQ benchmark is built from literature and labeled by whether GPT-4 and GPT-3.5 answer correctly.

Key results

• Fine-tuning InternLM2-7B with CQIA+CROP raises overall CROP benchmark accuracy from 0.368 to 0.768, and LLaMA3-8B from 0.348 to 0.754.
• Claude-3 Opus achieves the highest commercial overall accuracy (0.900) while GPT-3.5 scores 0.328 on the CROP benchmark.
• The paper reports an average improvement of 9.2% from CROP fine-tuning across open-source models, while Qwen1.5-7B gains are marginal (0.646 to 0.709 with CQIA+CROP).

Measured effects reported in the paper

Surprising findings

• Fine-tuning Qwen1.5-7B on CROP alone reduces accuracy on difficult questions (e.g., 0.302 to 0.202), and CQIA+CROP lowers difficult accuracy for Qwen1.5 and LLaMA3 combined training.
• GPT-4 scores 0.000 on difficult questions by construction, since difficulty labels are derived from GPT-4 and GPT-3.5 correctness.

Limitations

• Only rice and corn are covered; wheat, tuber crops and beans are absent.
• Text-only raw data; images and charts from papers are not used.
• Only 7-8B open models were fine-tuned; larger 70B/110B models were not tested.

Experimental setup

Datasets: CROP dataset CQIA Web of Science research papers professional books crop enterprise reports structured rice data

Benchmarks: CROP benchmark

Metrics: accuracy

Baselines: GPT-4 GPT-3.5 Claude-3 Qwen-max CQIA

Builds on: CAMEL CQIA CAMEL-style multi-agent role-play

Closest papers by meaning

• SuperGPQA: Scaling LLM Evaluation across 285 Graduate Disciplines (NeurIPS 2025 · similarity 0.88)
• AgMMU: A Comprehensive Agricultural Multimodal Understanding Benchmark (NeurIPS 2025 · similarity 0.88)
• AgroOmni: A Large-Scale Multi-view Agricultural Dataset for Cross-Scale Multimodal Reasoning (NeurIPS 2026 · similarity 0.86)
• What can large language models do for sustainable food? (ICML 2025 · similarity 0.86)
• MIRAGE: A Benchmark for Multimodal Information-Seeking and Reasoning in Agricultural Expert-Guided Conversations (NeurIPS 2025 · similarity 0.86)
• Unveiling the Secret Recipe: A Guide For Supervised Fine-Tuning Small LLMs (ICLR 2025 · similarity 0.85)