Topic: secret_language

A recent study shows that large language models perform better with certain languages and disadvantage others based on their training data composition. This highlights the need for more inclusive training datasets to ensure equitable performance across different language varieties.

arxiv.org↗