Hindi serves as the highest-resource Indic language in our cross-lingual transfer study, providing a comparative anchor between English and the low-resource languages Maithili and Bhojpuri. We fine-tune XLM-RoBERTa on 20,000 balanced Hindi sentences sampled from the iam-tsr/hindi-sentiments corpus (MIT License) and evaluate on a fixed test set of 501 sentences. English BERT in zero-shot achieves 35.33% (F1: 0.2263), consistent with its failure across all three Indic languages in this study. XLM-RoBERTa zero-shot recovers to 63.07% (F1: 0.6339), and fine-tuning yields 70.66% (F1: 0.7063). Notably, out-of-distribution evaluation on 30 manually written Hindi sentences produces 90.00% accuracy (F1: 0.9017) — the highest real-world generalisation score across all languages in this study, despite Hindi achieving the lowest in-distribution fine-tuned score. We attribute this to the translated training corpus producing a model that is less overfit to dataset-specific patterns. These results su…
