metadata
license: apache-2.0
This dataset helps small language models become better at instruction-following tasks, especially the ones on visual, byte-level modifications.
This dataset is meant to make small models more nuanced on their instruction-following skills.
It is mostly useful for byte-level models since a great part of it requires granularity bpe models usually lack.
The dataset is primarly dominated by visually charged domains, like ascii art, emojis and char-level puzzling multi-instruction texts.