Abstract
Abstract
Crop recommendation is an important decision-support task in agriculture, where multiple crops may be suitable under similar soil and environmental conditions. In this paper, we propose a hybrid framework that combines classical machine learning with a fine-tuned lightweight language model for soil-based crop recommendation. We first evaluate several classical machine learning models and a neural network baseline on tabular agricultural data. Our results show that XGBoost achieves the best predictive performance (0.487 Top-1 and 0.770 Top-3 accuracy across 48 crop classes), while a Multi-Layer Perceptron does not provide meaningful improvement over tree-based methods. Because strict single-label prediction is often inadequate in agricultural settings, we reformulate the task as a Top-K recommendation problem and show that Top-3 accuracy provides a more realistic measure of practical utility. To improve interpretability and usability, we further fine-tune a lightweight language model using LoRA to generate structured, farmer-friendly recommendations conditioned on machine learning predictions. Qualitative evaluation on held-out examples demonstrates that the fine-tuned model produces more coherent, grounded, and useful outputs than the base model. Overall, our results highlight the effectiveness of combining tabular machine learning for prediction with language models for explanation in agricultural decision-support systems.