PROJECT
The KRong Project, which provides LLM students with an understanding of the formation process of the Korean language.
Flagship Models
- KRongBERT
- An enhanced factorization‑based morphological LM published in IP&M (JCR 4.4 %), boosting token‑level F1 by 4.2 points over KoBERT.
- KRongDEFT
- A Korean Alpaca variant built via embedding‑layer quantization; delivers the same chat quality with 40 % fewer GPU hours (GitHub ★ 1.1 k).
- KRongLlama(Ongoing)
- Builds on KRong’s morphology‑aware embeddings to adapt the Llama architecture for high‑fidelity Korean instruction tuning.
- Action‑Concentrated Embedding (ACE)
- “This is your captain sign‑tokening,” selected for COLING 2024, augments action tokens with an 18 % accuracy gain.
Advanced Research Lab
- Omni³ LLM Lab
- Human–LLM synergy for Omni‑modal, Omni‑context‑aware, Omnipresent intelligence, funded as a Basic Research Laboratory project.
KRong-based Applied R&D Programs
Project 01
Standardizing Flood-Related Data(KISTI)
Powered by the KRongLLM, we converted heterogeneous, non‑standard flood‑related datasets from sensors, satellite imagery and municipal logs into a unified schema and production‑grade pipeline, providing the data backbone for high‑accuracy urban inundation forecasting and real‑time response systems.
Project 02
Multi‑Hazard Nowcasting & Prevention Tech (Nationwide Disaster‑Graph AGI)
Models the whole of Korea as a heterogeneous, multi‑layer graph, trains KRong LLMs on human‑centric real‑time disaster signals, and delivers an autonomous agent that anticipates compound hazards and initiates self‑directed response actions.
