The Ministry of Science and ICT (MSIT) of South Korea has officially announced the results of the second-round evaluation for the "Independent AI Foundation Model Project" (DOK-PA-MO). This initiative is a strategic governmental effort aimed at fostering domestic technological sovereignty in artificial intelligence by cultivating indigenous foundational model capabilities.

Following a rigorous assessment process, three consortia were selected as the final participants: LG AI Research, SK Telecom, and Upstage. These entities will receive state support to further develop their proprietary Large Language Models (LLMs) and integrated AI infrastructures. Conversely, Motif Technology failed to meet the evaluation criteria and was excluded from the final selection.

The core methodology of this project focuses on accelerating the development of domestic Foundation Models (FMs), which are defined as large-scale neural network architectures trained on vast, heterogeneous datasets using self-supervised learning techniques. The objective of the state-sponsored program is to enhance the performance metrics of these models, specifically optimizing for parameters such as perplexity ($PPL$) and computational efficiency.

The evaluation framework utilized by the MSIT involves assessing the models' architectural scalability, measured by the relationship between the number of parameters ($\theta$) and the training token count ($N$):
$$C \approx 6 \times N \times \theta$$
where $C$ represents the total floating-point operations (FLOPs) required for training. Participants were evaluated on their ability to optimize this scaling law to maximize performance within restricted computational budgets. Furthermore, the selection process emphasized the integration of Retrieval-Augmented Generation (RAG) capabilities and domain-specific fine-tuning strategies to ensure the models achieve high contextual accuracy ($A_{context}$) and minimize hallucination rates ($R_{hallucination}$), defined as:
$$R_{hallucination} = \frac{N_{incorrect\_claims}}{N_{total\_claims}}$$
Through this initiative, the MSIT seeks to strengthen the domestic AI ecosystem by promoting high-performance, cost-effective models capable of competing on a global scale.