metadata
language:
- en
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 is also useful for byte-level models since a great part of it requires granularity bpe models usually lack.
It followed human-curated decontamination and syntax-level dedup.
The final dataset contains around 205k examples and we'll continue expanding it.