Polite on the Surface, Wrong in Practice: A Curated Dataset for Fixing Honorific Failures in Multilingual Bangla Generation

Researchers introduced BLADE, a curated 4,196-pair Bangla dataset and benchmark for instruction tuning.

The resource focuses on fixing honorific, pragmatic, and structural errors in Bangla text generation.

Fine-tuning open-weight models like DeepSeek-8B and LLaMA-3.2-3B improved honorific alignment and output fidelity.

The work addresses a common multilingual AI failure mode: fluent but socially or culturally wrong text, especially in low-resource languages.