A team led by Md Nishat Raihan published BanglaShield on arXiv. The resource contains 10,245 Bengali prompts in three script forms, each labeled for safety. When tested on six current large language models, those models marked 47 % of unsafe native‑script prompts as safe, and 63 % when the text was romanized. BanglaShield, a compact fine‑tuned encoder, achieved 87.8 macro‑F1 on native‑script and 84.4 on romanized prompts, beating the best LLM by about 45 points while using roughly 130 × fewer parameters.
Why it matters
Developers of Bengali language applications can use BanglaShield to filter unsafe content more reliably than current LLM‑based classifiers.
