BanglaShield: Safety Classifier for Multimodal Bangla Text
October 20, 2026
BanglaShield provides a safety benchmark for native, Romanized, and code-mixed Bangla, revealing that LLM safety accuracy drops by 10.3 points when handling response-level safety in these formats.
HOW THIS AFFECTS YOU
●
researcherThe dataset highlights significant safety gaps in LLMs for low-resource, code-mixed languages.
●
policyStandard safety protocols may fail to protect users in non-English script environments.