We take open models like Llama and Qwen and retrain them on Bangla text. That includes teaching them to read Bangla script more efficiently and to follow instructions written in Bangla. The BanglaLlama family runs from 1B to 13B parameters, and every size is free to download on HuggingFace.
Safety training done in English does not survive the trip to Bengali. In BanglaSafe, our EMNLP 2026 Findings paper, we asked models the same harmful question several ways and changed nothing but the writing style. Written like a newspaper article, it got answered 17.5 points more often than written like a casual text message. We build the tests that catch this, and we sort the harms by what Bangladeshi law actually treats as a crime.
If you want to know whether a model is any good at Bangla, there is very little to test it with. We are building those tests: whether a model can spot political slant in news, whether it can work through a math problem in Bangla, and whether it gets better when you give it more time to think. The tools to run them are open too.
A model is only as good as the text it learns from, and for Bangla most of that text did not exist in usable form, so we made it. We collect Bangla news, translate instruction datasets into Bangla (Bangla-Alpaca and Bangla-Orca), and build math datasets. All of it is public.
We would rather the work end up in something people use than sit in a paper. Drishtikon, an app that lets Bangladeshi readers compare how different outlets covered the same story, is built on this lab's work.
We wrote 879 harmful questions in Bengali rather than translating them from English, covering 17 kinds of harm that Bangladeshi law treats as a crime. Keep the question the same and change only how it is written, and models answer 17.5 points more often. Measured over 15,822 answers from 18 models, each graded on a four-point scale that was checked against human graders.
arXiv · Project page · Code · Dataset
Introduces two Bangla training sets, Bangla-Alpaca with 52,000 examples and Bangla-Orca with 172,000, and releases 5 BanglaLlama models built with them.
BanglaBias, 200 Bangla news articles, each labeled as leaning towards the government, critical of it, or neutral. We tested 28 language models on how well they could tell the difference.
Everything below is free to download from HuggingFace: 31 models and 8 datasets.
Llama 2, 3, 3.1 and 3.2 retrained on Bangla, from 1B to 13B parameters. Each size comes in a plain version and one tuned to follow instructions.
Models given more time to think before they answer, adapted for Bengali. Built on Qwen-2.5 at 3B and 32B, and on QWQ-32B.
A safety test set in Bengali. 17 kinds of harm taken from Bangladeshi law, with each question written in five different styles. From the EMNLP 2026 Findings paper.
Bangla questions paired with answers, used to teach a model to follow instructions. Alpaca and Orca combined.
The Alpaca instruction set translated into Bangla.
Math problems in Bangla, with the solutions worked out.
drishtikon.life · Parent: Perspectivity
We work with students, researchers, and engineers, and we are glad to hear from anyone who wants to help build Bangla language technology.