Benchmarking Intent Classification in Extremely Low-Resource Indo-Aryan Dialects: An Empirical Study on Dhatki Agricultural Telemetry

Natural Language Processing (NLP) models achieve state-of-the-art benchmarks in high-resource global languages, yet regional vernacular dialects spoken by millions of marginalized pastoralists remain entirely excluded from digital conversational interfaces. Dhatki, an Indo-Aryan transitional language spoken by over 2.5 million inhabitants across the Thar Desert of southeastern Sindh, Pakistan, lacks standardized tokenizers, orthographic corpora, and annotated evaluation benchmarks. In this technical report, we introduce the first domain-specific, curated Intent Classification Dataset for Dhatki agricultural telemetry (comprising 1,000 verified acoustic and textual utterances spanning 10 operational agricultural classes, hosted openly on Hugging Face). We empirically benchmark baseline lexical models against multilingual transformer representations (mBERT, IndicBERT, XLM-RoBERTa, and a proposed domain-specific Dhatki-DistilBERT) under severe resource constraints. Our experimental findings demonstrate that custom subword Byte-Pair Encoding (BPE) tuned to phonetic Dhatki transcriptions achieves a Micro-F1 score of 89.4% with an edge latency of 14.2 ms, proving the viability of low-latency, offline voice interfaces for low-resource desert agriculture. Open-Access Resources: • Hugging Face Dataset: https://huggingface.co/datasets/lokeshsonara/dhatki-agri-intents • Live Research Portal: https://agri-idas.tech/