📄 Paper
We introduce KMMLU-Redux, a reconstructed version of the existing KMMLU, comprising 2,587 problems from Korean National Technical Qualification (KNTQ) exams. We identified several critical issues in the KMMLU, including leaked answers, lack of clarity, ill-posed questions, notation errors, and contamination risks. To address these problems, we conducted rigorous manual examination and decontaminated the dataset from the pre-training corpus to prevent potential… See the full description on the dataset page:
https://huggingface.co/datasets/LGAI-EXAONE/KMMLU-Redux.
