rewrite

This dataset helps small language models become better at instruction-following tasks, especially the ones requiring semantic or char-level modifications.

Baseline for continuous high-quality updates.

This dataset is meant to make small models more nuanced on their instruction-following skills.

It followed human-curated decontamination and syntax-level dedup.

The final dataset contains around 205k examples and we'll continue expanding it.