Researchers have developed KrishokChat, a new dataset and benchmark designed to improve agricultural advisory services in Bengali for low-resource settings. The dataset includes over 145,000 question-answering pairs, grounded in 129 agricultural manuals and featuring verified citation provenance for all information. A Farmer Benchmark of 1,001 real-world farmer queries was also created to evaluate model performance. Initial testing with Gemma-4-E2B showed that while KrishokChat improves structured formatting, models still face challenges with precise chemical dosage generalization, indicating its primary utility for retrieval-augmented generation. AI
IMPACT Enhances AI capabilities for agricultural advice in Bengali, particularly for retrieval-augmented generation systems.
RANK_REASON Publication of a new dataset and benchmark for a specific domain (agricultural advisory) in a low-resource language. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
