Projects
Research projects in NLP, Information Retrieval, and RAG.
WBL: World Best LLM Project HuggingFace
- Led the data team and established a query clarity evaluation framework using engineered GPT-5.2 prompts, training a Qwen3-4B model as a proprietary tagger to selectively curate datasets with diverse clarity levels.
- Established a robust response filtering pipeline for off-policy SFT data by ensembling three distinct reward models, applying score fusion techniques to accurately evaluate and retain high-quality responses.
- Curated and refined large-scale alignment samples by integrating the score fusion pipeline with rigorous LLM-as-a-judge and Code Execution metrics to effectively eliminate repetitive and low-quality responses.
KURE: Korea University Retrieval Models GitHub | HF | API
- Led the lab’s flagship Korean-English retrieval project, built the MTEB-Ko-Retrieval Leaderboard, and open-sourced the model family and training/evaluation code, reaching 250+ GitHub stars.
- Dense & Multi-Vector: Developed KURE-v1, KURE-v2, and KURE-v3. KURE-v1 and KURE-v2 both ranked 1st on the leaderboard (as of Aug. 2025 and Oct. 2026, respectively). KURE-v1 surpassed 3M downloads, KURE-v2, a multi-vector model, outperforms 8B dense models while being 50× smaller, and KURE-v3 outperforms Qwen3-Embedding-8B while being 7× smaller and is served through a public API.
- Sparse: Developed KURE-Sparse-{base, nano}, achieving state-of-the-art Korean sparse retrieval performance overall and under 0.5B parameters, respectively. The base model also outperforms 8B dense models.
- Reranker: Developed KURE-Reranker-{base, nano}, achieving state-of-the-art Korean reranking performance under 4B and 1B parameters, respectively. The 149M nano delivers 19.5× the throughput of Qwen3-Reranker-4B with only a 1.5-point gap in nDCG@10.
ReviewSearch Blog | GitHub | Demo
- Built and deployed a search engine over 205,988 real peer reviews and rebuttals for agents, so that a rebuttal draft is grounded in how authors actually answered the same concern.
- Trained a hybrid retriever for the domain: a fine-tuned dense encoder and a sparse encoder fused with Reciprocal Rank Fusion.
- Shipped it as the grounding layer for
RebuttalDraftin therebuttal-skillsagent plugin.
URACLE-Korea University Collaborative Research
- Trained a Korean-English cross-lingual retriever and analyzed language-pair trade-offs. Used model merging to recover mono-lingual retrieval while retaining CLIR gains. [Paper]
KT-Korea University Collaborative Research (Korean Legal LLM)
- Developed an end-to-end training recipe for a Korean legal-domain LLM. Defined real-world judicial tasks, curated expert-written and synthetic alignment datasets, and optimized training strategies to preserve general capabilities.
- This research directly contributed to KT winning a $10.42 million contract to build an AI platform for the South Korean Supreme Court. [News]
- Published the training methodologies and data pipelines as LEGALMIDM at the ICLR 2026 Data-FM Workshop.