---
license: apache-2.0
---
[GitHub](https://github.com/pds-dpo/pds-dpo) | [arXiv](https://arxiv.org/abs/2412.17417)

This dataset is proposed for post-training where we use a very small amount of data to get exactly the behaviors we want. Do note that small amounts of the right kinds of behavior (safety, instruction-following, style) can also make a difference, although there is a long-tail benefits from more data.