Earlier Yi paper indicates it's trained with less than 25% Chinese dataset, contrast to GPT-3 which was 93% English[1][2]. Is that a bug or could there be somet

Earlier Yi paper indicates it's trained with less than 25% Chinese dataset, contrast to GPT-3 which was 93% English[1][2]. Is that a bug or could there be something inherent to current LLM architecture - like dataset must be 90%+ English to not fall apart? 1: https://arxiv.org/html/2403.04652v1 https://arxiv.org/html/2403.04652v1 2: https://github.com/openai/gpt-3/blob/master/dataset_statistics/languages_by_word_count.csv https://github.com/openai/gpt-3/blob/master/dataset_statisti...