Rebuild of the keural-SFT mixed EN/KO instruction dataset, reprocessed from the original
public sources with a fixed cleaning stage.
⚠️ Why this rebuild exists: the original mkd-chanwoo/keural-SFT was produced with a
cleaner rule that collapsed consecutive spaces, which flattened all code-block indentation
to one space — code samples were syntactically broken (root cause of a HumanEval
regression). The corruption is not recoverable from… See the full description on the dataset page:
https://huggingface.co/datasets/Mkd-Yonas/keural-SFT-rebuilt.
