maius/OpenCharacterTraining-data-misalignment

Open Character Training

Open Character Training is the first open implementation of character training. For more information, read our paper!

This repository includes all training data generated for our paper for the misalignment persona only. See Section 2 for details of DPO, self-reflection, and self-interaction. It is intended for further research / replication.

Please get in touch for further details.

License

This dataset follows the same license used in LIMA, which is a source of several prompts.
"If the source data of LIMA has a stricter license than CC BY-NC-SA, the LIMA dataset follows the same. Otherwise, it follows the CC BY-NC-SA license."

Citation

@misc{maiya2025opencharactertrainingshaping,
      title={Open Character Training: Shaping the Persona of AI Assistants through Constitutional AI}, 
      author={Sharan Maiya and Henning Bartsch and Nathan Lambert and Evan Hubinger},
      year={2025},
      eprint={2511.01689},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2511.01689}, 
}