during multi-language model training, accuracy of English/numbers is worse than Chinese characters, how to improve?