Tag
This paper releases GLAN-QnA-KR, a 303,581-row Korean instruction-QA corpus generated via a seedless taxonomy-driven pipeline using Phi-3.5-MoE-instruct, with near-zero duplication and low benchmark contamination.