m-ArenaHard
收藏资源简介:
m-ArenaHard数据集是由Cohere For AI创建的多语言评估数据集,包含500条翻译自英语的提示,覆盖23种语言。该数据集旨在通过Google Translate API翻译原始的Arena-Hard-Auto数据集,以评估多语言模型在不同语言环境下的表现。创建过程中,数据集利用了多种创新方法,如多语言数据套利和模型合并,以提高模型的多语言性能。m-ArenaHard数据集主要应用于多语言AI模型的评估和优化,旨在解决多语言模型在不同语言间性能不均衡的问题。
The m-ArenaHard dataset is a multilingual evaluation dataset developed by Cohere For AI, which contains 500 prompts translated from English and covers 23 languages. This dataset is constructed by translating the original Arena-Hard-Auto dataset via the Google Translate API, aiming to evaluate the performance of multilingual AI models across diverse linguistic contexts. During its creation, multiple innovative methods including multilingual data arbitrage and model merging were adopted to enhance the multilingual capabilities of AI models. The m-ArenaHard dataset is mainly applied to the evaluation and optimization of multilingual AI models, with the goal of addressing the issue of imbalanced performance of such models across different languages.

- 1Aya Expanse: Combining Research Breakthroughs for a New Multilingual FrontierCohere For AI · 2024年



