Extending English Large Language Models to New Languages

Large language models (LLMs) demonstrate remarkable proficiency across a broad spectrum of natural language processing (NLP) tasks exhibiting knowledge, common-sense and reasoning skills. While these LLMs perform well for English, their performance on non-English languages, particularly low-resources ones, is significantly lower. This highlights a significant challenge in extending their benefits to non-English languages. In this talk, I will discuss the technical challenges involved and various approaches in recent research to extend English LLMs to other languages and close the gap with English language performance. The talk will cover a broad range of topics from the entire LLM stack – from tokenizer to instruction tuning – that are relevant to making LLMs multilingual. The talk will also cover of our recent efforts in this space and in enabling Indian languages with the capabilities that LLMs provide.

Github Page: https://github.com/anoopkunchukuttan/...