cmu-lti/osim-post-training

SOUL

This CMU-LTI mirror hosts the post-training data used for ODYSSIM releases. It mirrors the original sunweiwei/Soul dataset layout under the CMU-LTI organization.

SOUL is the data suite for human behavior simulation used in Reinforcing Human Behavior Simulation via Verbal Feedback (DITTO), spanning conversation, social simulation, social cognition, role-play, and human-centric evaluation.

• —📄 Paper: https://arxiv.org/abs/2605.20506
• —💻 Code: https://github.com/sunnweiwei/OdysSim
• —🤖 Model: https://huggingface.co/sunweiwei/Ditto-8B

Structure

Files are kept in their original format, bound to our training/eval framework, and are organized into two folders:

• —train/ — RL training data, one parquet per task.
• —test/ — evaluation data, one parquet per task.

Because each task's extra_info schema differs across tasks and between train/eval, each task+split is exposed as its own config (<task>_train, <task>_test) so it loads cleanly:

from datasets import load_dataset
train = load_dataset("cmu-lti/osim-post-training", "userllm_train", split="train")
test  = load_dataset("cmu-lti/osim-post-training", "userllm_test",  split="test")

You can also read any parquet directly, e.g. train/userllm_rl_train.parquet / test/userllm_val.parquet.

Citation

@article{sun2026ditto,
  title         = {Reinforcing Human Behavior Simulation via Verbal Feedback},
  author        = {Sun, Weiwei and Zhou, Xuhui and Liu, Jiarui and Du, Weihua and Sun, Haojia and Xie, Yiqing and Ma, Qianou and Chen, Sihao and Wan, Mengting and Yang, Longqi and Zhou, Pei and Wu, Sherry and Welleck, Sean and Neubig, Graham and Yang, Yiming and Sap, Maarten},
  year          = {2026},
  eprint        = {2605.20506},
  archivePrefix = {arXiv},
  url           = {http://arxiv.org/abs/2605.20506}
}