South Korean tech company Naver has officially introduced HyperCLOVA X, a new suite of large language models (LLMs) designed to enhance cross-lingual reasoning capabilities, particularly in Asian languages. The company claims these models outperform existing models in understanding and processing multiple languages, potentially aiding regions in developing their own sovereign LLMs. An English-language technical report detailing HyperCLOVA X's capabilities has been published on the arXiv open access journal, emphasizing its proficiency not only in English and Korean but also in a range of other languages.
Advanced Multilingual Capabilities
HyperCLOVA X has been pre-trained on a diverse dataset that includes Korean, English, and code segments, with a significant portion of the data being in Korean. This approach aims to refine the model's performance in Naver's primary language while also ensuring it can handle other languages effectively. The technical report highlights HyperCLOVA X's state-of-the-art performance in machine translation between Korean and other languages, such as Japanese and Chinese, which were not specifically targeted during its training. Furthermore, the model demonstrates exceptional cross-lingual transfer skills, particularly between Korean and English, suggesting that instruction-tuning in one language can activate instruction-following capabilities in another.
A Step Towards Sovereign AI
The development of HyperCLOVA X aligns with the growing recognition of sovereign AI as a critical national asset. This concept, supported by companies like Nvidia, aims to ensure data security and reduce reliance on foreign AI technologies. Naver's initiative addresses the overrepresentation of English and North American cultures in existing LLMs' training data, which often limits their effectiveness in processing non-English languages and understanding their unique cultural and linguistic nuances. By focusing on multilingual and multimodal capabilities, HyperCLOVA X aims to provide a more inclusive and versatile AI model that can cater to the specific needs of various regions and countries.
Naver's commitment to enhancing HyperCLOVA X's functionality extends to exploring multimodality and integrating external tools and APIs. This effort is expected to enable the model to process and integrate diverse types of data seamlessly, further broadening its applicability and effectiveness in real-world scenarios.
