During the annual exchange program between the STPI and the Korea Institute of Science and Technology Information (KISTI), the discussions on the first day focused on KISTI's data integration, foundational data infrastructure, and how KISTI applies AI to integrate data from various sources. The dialogue included topics such as KISTI's training of large language models (LLMs) and how these models support Korean researchers in conducting academic research more efficiently.

(Above) STPI researchers photo with the Director of National S&T Data Division (center)

The Large-Scale AI Research team of KISTI introduced its recent advancements in training language models.  Notably, KISTI has trained a Korean LLM based on Llama 2 in 2023.  In 2024, KISTI successfully developed an academic AI model “KONI” by using Llama 3 to enhance its performance. This model aims to provide researchers with effective academic tools. The team trained the model using open Korean datasets and re-trained it by collecting data from multiple sources. They also employed “Retrieval-Augmented Generation (RAG)” technology to ensure that the model generates fact-based responses. Additionally, KISTI even improved the model's accuracy validation through the integration of a RAG pilot service.

The visit agenda also included the “KISTI Tour”, which showcased KISTI’s primary services, their visionary blueprint of using data to change the world, and the evolution of how KISTI developed supercomputers.

(Above) Dr. Jayhoon Kim from Division of National S&T Data introduced KISTI’s Digital Curation Model 

(Above) KISTI services introduced during the KISTI Tour

In addition to visiting KISTI's headquarters in Daejeon, the delegation also visited KISTI’s Seoul branch as well as the Korea Institute of Science & Technology Evaluation and Planning (KISTEP).  At KISTI’s Seoul branch, both parties exchanged insights about KISTI’s 2 renowned platforms, AccessON and ScienceON.  The ScienceON team introduced how they take advantage of LLM techniques to provide interactive models on the platform and how they use AI technologies for literature review.

The primary purpose of the visit to KISTEP was to discuss the research methodologies applied in their Technology Level Evaluation Report.  Through in-depth interviews, it’s learned that KISTEP’s methodology combines patent and literature analysis with the Delphi method.  For quantitative analysis, KISTEP used 15 different indicators to assess the level of technological development in various countries. These quantitative results serve as the basis for expert evaluations, and the experts assess each country's level of advancement in these technologies through the Delphi method.  This analysis helps the Korean government understand the national level of technological development and identify key technology strengths in order to facilitate the development of core technologies in Korea.

 (Above) STPI delegates photo with KISTEP research team

The annual exchange program between STPI and KISTI not only deepened the understanding of open data, resource integration, LLMs, and the design of research indicator in Korea, but also provided valuable insights into the preferences of local researchers, research trends, innovative technology applications and supporting tools, data processing, as well as natural language processing technologies.  These practical experiences offer valuable references for STPI researchers in the planning of research process and development of supporting tools.