language:
- en
This is one of the core datasets used with palmer models alongside other private datasets.
This one was created as a baseline for continuous high-quality updates. This is an internal dataset meant to make small models more nuanced on their instruction-following skills. The original dataset comprises a total of 1 million examples outsourced from a diverse set of sources, including additional synthetic generation. It followed human-curated decontamination, after this, it was syntax-level deduped The final dataset contained around 195k examples and we continued expanding it with humanly curated batches of approximately 1k each during several weeks.
This dataset is still expanding and being continuously refined over time. Keep in mind that this is not the only dataset used for palmer models and is not cleaned either. So use with precaution.