Tag
This paper introduces BLADE, a culturally aligned instruction-tuning dataset of 4,196 interaction pairs for fixing honorific failures and pragmatic gaps in multilingual Bangla generation. Fine-tuning models like DeepSeek-8B and LLaMA-3.2-3B on this dataset yields substantial improvements in structural fidelity and honorific alignment.