Final Year Project

Project Title

ALIF - A Pretrained Urdu Generative Model

Description

Alif is a language model for Urdu which understands cultural nuances and contexts without being computationally expensive. Large Multilingual Language Models are mostly trained for languages like English, French, German. For underrepresented languages like Urdu, these models often produce ill-nuanced responses. State-of-the-art LLMs like GPT perform well to an extent in Urdu, but too computationally. Environmentally and financially, they are expensive to use. For local governments using these models could incur a huge training, and end point usage cost. Organizations like OpenAI cannot be trusted with sensitive data of banks and governments which is why using API is not a viable option. Therefore, we are working to build a Small Language Model which is significantly smaller in size than the large multilingual models but gives similar performance for Urdu language.